<?xml version="1.0" encoding="UTF-8"?><rss version="2.0" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/"><channel><title>Security - Soursen</title><description>Security covers vulnerabilities and the software supply chain, checked against the advisory or the incident report that records what failed.</description><link>https://soursen.org/</link><language>en-gb</language><atom:link href="https://soursen.org/beat/security/rss.xml" rel="self" type="application/rss+xml"/><item><title>Claude Code’s “Don’t Ask Again” Button Has Put 2,539 Secrets on Public GitHub</title><link>https://soursen.org/articles/claude-codes-dont-ask-again-button-has-put-2-539-secrets-on-public/</link><guid isPermaLink="true">https://soursen.org/articles/claude-codes-dont-ask-again-button-has-put-2-539-secrets-on-public/</guid><description>A permission shortcut saves the exact command you approved, password included, into a project file that often gets committed. The bigger lesson is that AI coding agents have created new places for secrets to hide, and the usual scanners are not looking there.</description><pubDate>Sun, 08 Nov 2026 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;On November 5, 2026, researchers Shmulik Cohen and Matan Giladi at the application security company Apiiro published a scan of public GitHub, and the numbers are uncomfortable for anyone who uses Claude Code. They found 104,185 public copies of a Claude Code file called .claude/settings.local.json. 1,122 of them, about 1.1%, held a real secret. Counting older versions in git history, the shared project settings file and saved agent chat logs, the total came to 2,539 distinct secrets across 1,994 public repositories, owned by 63 organizations. I want to walk through how a single click turns into a stored password, why deleting the file does not fix it, why standard secret scanners missed hundreds of these and what you should check in your own repositories today.&lt;/p&gt;
&lt;h2 id=&quot;how-1-click-turns-into-a-stored-password&quot;&gt;How 1 click turns into a stored password&lt;/h2&gt;
&lt;p&gt;Claude Code asks for permission before it runs a shell command. One of the options is “Yes, and don’t ask again.” Pick that, and Claude Code saves an allow rule so it can run the same command later without asking. According to &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://code.claude.com/docs/en/settings&quot;&gt;Claude Code’s settings documentation&lt;/a&gt;, those personal rules live in .claude/settings.local.json inside the project folder.&lt;/p&gt;
&lt;p&gt;The rule is the command itself, saved word for word. If the command you approved was a database login with the password typed inline, or a curl request with an authorization header, the password or token goes into the file with it. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://thehackernews.com/2026/11/claude-codes-dont-ask-again-file.html&quot;&gt;The Hacker News reported&lt;/a&gt; that Apiiro reproduced the behavior with Opus 5.5 and Fable 5.1, and that of 16 coding agents it compared, Claude Code was the only one confirmed to save the full approved command in a project file by default.&lt;/p&gt;
&lt;p&gt;Most developers have learned to keep .env files out of git. Very few have ever opened .claude/settings.local.json, and the name suggests it stays local. It only stays local if git is told to ignore it.&lt;/p&gt;
&lt;h2 id=&quot;why-passwords-end-up-inside-commands&quot;&gt;Why passwords end up inside commands&lt;/h2&gt;
&lt;p&gt;Typing a secret into a command is a normal shortcut. A developer debugging a database connects with the password set inline because it is faster than configuring a credentials file. Someone testing an API pastes a bearer token into a curl header. Someone deploying to a server passes an SSH password to a login tool. On a plain terminal, that command lands in shell history on 1 laptop, which is a mild risk.&lt;/p&gt;
&lt;p&gt;An agent changes the math. Claude Code proposes the command, the developer approves it, and a single click on “don’t ask again” turns that throwaway line into a saved configuration entry inside the project folder. Lakera’s researchers made the same point after their npm study, noting that developers approve large numbers of shell commands during routine work and that a token in an authorization header can be saved permanently the moment someone picks persistent approval over a one time run.&lt;/p&gt;
&lt;h2 id=&quot;what-the-file-is-supposed-to-do&quot;&gt;What the file is supposed to do&lt;/h2&gt;
&lt;p&gt;Anthropic designed the file to stay private. Claude Code’s documentation describes .claude/settings.local.json as the place for personal settings that are not checked into source control, separate from .claude/settings.json, which teams share and commit on purpose. The documentation says that when Claude Code creates the local file, it configures git to ignore it.&lt;/p&gt;
&lt;p&gt;That protection is real, and it has gaps. Apiiro found that on the first write inside a git repository, Claude Code adds the file to the developer’s global git excludes, a setting that lives on that 1 computer for that 1 user. It does nothing for a file that was already committed. It does not travel with the project, because Claude Code never adds the file to the project’s own .gitignore. A teammate on another machine, a CI job, a fresh clone or a developer who created the file by hand gets no protection at all. As &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://itwire.com/business-it-news/security/claude-code-approval-files-found-exposing-company-secrets-on-github-says-apirro-researchers&quot;&gt;iTWire noted&lt;/a&gt;, the global exclude is a real protection. It just covers a narrower set of cases than most people assume.&lt;/p&gt;
&lt;h2 id=&quot;what-apiiro-found&quot;&gt;What Apiiro found&lt;/h2&gt;
&lt;p&gt;Apiiro’s scan covered 104,185 copies of the file in 103,617 public repositories. The 1,122 copies with a real secret held 1,379 distinct values. Older versions of the file that only survive in git history added 734 more secrets across 509 repositories. The shared .claude/settings.json file had secrets in 298 of 76,011 copies, mostly inside allow rules, and saved agent chat logs added more.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-01.n2T-cZpN_Z162d8W.webp&quot; alt=&quot;Secrets tied to Claude Code settings files, by where Apiiro found them.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1488&quot; height=&quot;838&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Secrets tied to Claude Code settings files, by where Apiiro found them. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://thehackernews.com/2026/11/claude-codes-dont-ask-again-file.html&quot;&gt;Apiiro via The Hacker News, Nov 2026&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;p&gt;The secrets were the kind that open doors. Apiiro counted 379 database passwords, 43 SSH logins for public servers, 42 tokens for GitHub, GitLab and npm, and 23 AWS keys, &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://news.lavx.hu/article/claude-code-s-don-t-ask-again-setting-exposes-secrets-in-public-github-repositories&quot;&gt;according to LavX News&lt;/a&gt;. Among the named examples were a database vendor’s own admin level cloud key, a production Azure Postgres password, a Square production token and the owner password of a medical records database, &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://daily.dev/posts/claude-code-s-don-t-ask-again-exposes-passwords-in-1-122-public-github-files-5ing6vaqe&quot;&gt;per daily.dev’s summary&lt;/a&gt;.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-02.D68CTcTv_Z2cj4Ge.webp&quot; alt=&quot;The secret types Apiiro published counts for. Database passwords dominate.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1488&quot; height=&quot;838&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;The secret types Apiiro published counts for. Database passwords dominate. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://news.lavx.hu/article/claude-code-s-don-t-ask-again-setting-exposes-secrets-in-public-github-repositories&quot;&gt;Apiiro via The Hacker News and LavX News, Nov 2026&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;p&gt;The 63 owning organizations were mostly software companies of all sizes. Apiiro did not test whether the secrets still worked. It notified affected organizations, and 1 company confirmed that its key had been live.&lt;/p&gt;
&lt;h2 id=&quot;what-these-secrets-give-an-attacker&quot;&gt;What these secrets give an attacker&lt;/h2&gt;
&lt;p&gt;The categories matter more than the total. A database password paired with a hostname that answers on the public internet is direct access to the data, with no further hacking needed. An SSH login for a public server is a shell on that server. Apiiro’s advice to security teams is to deal with those pairs first, since a credential plus a reachable host is the shortest path to damage.&lt;/p&gt;
&lt;p&gt;Tokens for GitHub, GitLab and npm reach further. They can let someone push code or publish a package version under the owner’s name, which turns 1 leaked token into a risk for everyone downstream who installs that package. The Shai-Hulud worm in September 2025 showed how fast that kind of supply chain abuse spreads once tokens are in the wrong hands. Cloud keys such as the 23 AWS keys Apiiro found can be used to spin up computing resources and run up a bill, or to read whatever storage the key can reach.&lt;/p&gt;
&lt;h2 id=&quot;deleting-the-file-does-not-undo-the-leak&quot;&gt;Deleting the file does not undo the leak&lt;/h2&gt;
&lt;p&gt;The obvious fix is to delete the file. It does not work, and Apiiro’s numbers show why. Git keeps every committed version, so the 734 secrets that exist only in older versions are still readable by anyone who browses the history. Forks copy that history too, and Apiiro found 211 forks of 74 repositories carrying the same secrets.&lt;/p&gt;
&lt;p&gt;The files also sat in the open for a long time. When Apiiro last checked before publishing, every file with a secret was still on its repository’s default branch, and in half of those repositories it had been there for at least 217 days, The Hacker News reported. Copies outside GitHub matter as well. The same file can end up inside npm packages, Docker images, zip archives and synced folders.&lt;/p&gt;
&lt;p&gt;The only real fix for an exposed secret is to rotate it. Change the password, revoke the token, issue a new key, then clean up the repository.&lt;/p&gt;
&lt;h2 id=&quot;why-secret-scanners-missed-215-files&quot;&gt;Why secret scanners missed 215 files&lt;/h2&gt;
&lt;p&gt;Apiiro ran 2 popular open source scanners, gitleaks and TruffleHog, over every one of the 1,122 files, without live verification. Gitleaks flagged 564 of them. TruffleHog flagged 643. Neither flagged 215.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-03.DogVKk6n_zrr0t.webp&quot; alt=&quot;How many of the 1,122 files with real secrets each scanner flagged.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1488&quot; height=&quot;838&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;How many of the 1,122 files with real secrets each scanner flagged. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://thehackernews.com/2026/11/claude-codes-dont-ask-again-file.html&quot;&gt;Apiiro via The Hacker News, Nov 2026&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;p&gt;The misses follow a pattern. Most involved passwords typed inside commands, such as a value set in a Postgres password variable or passed straight to a MySQL or sshpass login. Scanners are tuned to recognize known key formats, like an AWS key or a GitHub token with a fixed prefix. A plain password sitting in the middle of a command string looks like ordinary text. That is the exact shape of secret the “don’t ask again” rule captures.&lt;/p&gt;
&lt;h2 id=&quot;claude-often-commits-the-file-itself&quot;&gt;Claude often commits the file itself&lt;/h2&gt;
&lt;p&gt;Apiiro traced the history of 1,105 repositories and found that in 539 of them, a commit listing Claude as a coauthor had changed the settings file. That suggests the agent itself is often the one adding the file to git, as part of a routine commit of everything in the folder.&lt;/p&gt;
&lt;p&gt;That lines up with earlier research from GitGuardian. Its &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://blog.gitguardian.com/the-state-of-secrets-sprawl-2026-pr&quot;&gt;State of Secrets Sprawl 2026 report&lt;/a&gt;, published in March, found that commits coauthored with Claude Code leaked secrets at about 3.2%, roughly twice the baseline across public GitHub. GitGuardian’s Dwayne McDaniel put the human only baseline at 1.5% in an &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://tfir.io/ai-code-secret-sprawl-gitguardian/&quot;&gt;interview with TFiR&lt;/a&gt;.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-04.PuNuSkyI_1sXPTl.webp&quot; alt=&quot;Leak rates from 3 separate studies. Each counts something different, so treat this as context.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1488&quot; height=&quot;838&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Leak rates from 3 separate studies. Each counts something different, so treat this as context. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://blog.gitguardian.com/the-state-of-secrets-sprawl-2026-pr&quot;&gt;Apiiro, GitGuardian, Lakera, 2026&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;p&gt;I would be careful reading the GitGuardian figure as proof that the AI writes worse code. Developers still decide what to approve, accept and push. What the data does show is that an agent committing quickly and often will commit whatever is sitting in the folder, including files a human might have paused over.&lt;/p&gt;
&lt;h2 id=&quot;the-bug-reports-that-a-bot-closed&quot;&gt;The bug reports that a bot closed&lt;/h2&gt;
&lt;p&gt;None of this is new to Anthropic. A bug report filed on GitHub on August 11, 2025, as &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://github.com/anthropics/claude-code/issues/5544&quot;&gt;issue #5544 in the Claude Code repository&lt;/a&gt;, said approved commands with API keys and tokens were being stored word for word. According to Apiiro, an automated bot closed it as “not planned” on January 3 after 60 days without activity. 4 related reports were closed by the same bot, the most recent the day before The Hacker News story ran, and Apiiro said none of them received a reply from Anthropic.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-05.BE39UJoX_Zb33ge.webp&quot; alt=&quot;Public warnings about secrets in Claude Code settings files since August 2025.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1488&quot; height=&quot;838&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Public warnings about secrets in Claude Code settings files since August 2025. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://thehackernews.com/2026/11/claude-codes-dont-ask-again-file.html&quot;&gt;Apiiro via The Hacker News, ironPeak, GitGuardian, Lakera&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;p&gt;Other users reported the same thing. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://claudeissues.com/issue/13106-bug-sensitive-tokens-stored-in-claude-settings-local-json-and-not-ignored-by-git&quot;&gt;Another GitHub report, issue #13106&lt;/a&gt;, described a developer whose key was flagged as leaked in a public repository, only to find the file holding that key and several other tokens. The reporter asked Claude Code to add the .claude folder to .gitignore automatically.&lt;/p&gt;
&lt;p&gt;In February 2026, Niels Hofmans of the security consultancy &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://ironpeak.be/blog/leaking-secrets-from-the-claud/&quot;&gt;ironPeak&lt;/a&gt; wrote about committing his own settings file by accident, with database passwords and API keys from approved commands still sitting in his git history. He released an open source scanner called claudleak that searches GitHub for AI tool folders such as .claude, .cursor and .continue and runs TruffleHog over them, history included. From a sample of 100 repositories he estimated that about 2.4% of repositories with AI tool folders hold sensitive information somewhere in their history. I could not find a public statement from Anthropic responding to the Apiiro research at the time of writing.&lt;/p&gt;
&lt;h2 id=&quot;the-claude-folder-is-security-sensitive-in-both-directions&quot;&gt;The .claude folder is security sensitive in both directions&lt;/h2&gt;
&lt;p&gt;The local settings file leaks secrets outward. The shared settings file can be a way in. In January 2026, Anthropic patched CVE-2026–21852, reported by Check Point Research, in which a repository’s .claude/settings.json could redirect Claude Code’s API traffic to another address before the user saw a trust prompt, the fix shipping in version 2.0.65, &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.backslash.security/blog/your-claude-code-token-might-end-somewhere-you-dont-expect&quot;&gt;according to Backslash Security&lt;/a&gt;. Backslash recommends treating .claude/settings.json as executable content during code review. Put the 2 findings together and the whole .claude folder deserves the same scrutiny as build scripts and secrets files.&lt;/p&gt;
&lt;h2 id=&quot;what-could-change-in-claude-code&quot;&gt;What could change in Claude Code&lt;/h2&gt;
&lt;p&gt;The people who reported the problem have suggested fixes. The author of issue #13106 asked Claude Code to add the .claude folder, or at least the local settings file, to the project’s .gitignore automatically when it creates it. Apiiro’s write up points to the gap between the global exclude and the project’s own .gitignore as the reason the current protection fails. 1 reader responding to the Lakera research on &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://bdtechtalks.substack.com/p/the-silent-threat-of-claude-code&quot;&gt;TechTalks&lt;/a&gt; suggested storing approved commands in a separate file that teams could ignore on its own.&lt;/p&gt;
&lt;p&gt;I think the most useful change would be the simplest one, which is warning the user, or masking the value, when a command about to be saved as a permanent rule contains something that looks like a password or token. Until something like that ships, the fix sits with developers and security teams.&lt;/p&gt;
&lt;h2 id=&quot;this-keeps-happening-across-ai-tools&quot;&gt;This keeps happening across AI tools&lt;/h2&gt;
&lt;p&gt;The settings file is 1 example of a wider pattern. In April 2026, researchers at Lakera, now part of Check Point, monitored about 46,500 npm packages and found 428 that shipped a .claude/settings.local.json file. 33 files across 30 packages contained credentials, roughly 1 in 13, including npm tokens, GitHub personal access tokens and production bearer tokens, &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://itbrief.co.uk/story/claude-code-can-leak-secrets-in-public-npm-packages&quot;&gt;as IT Brief reported&lt;/a&gt;. Published npm packages are effectively permanent, so every one of those credentials had to be treated as compromised.&lt;/p&gt;
&lt;p&gt;Steve Giguere, principal AI security advocate at Check Point, said at the time that AI coding assistants are introducing entirely new ways for secrets to be created, stored and exposed, &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://securitybrief.co.nz/story/claude-code-can-leak-secrets-in-public-npm-packages&quot;&gt;per SecurityBrief&lt;/a&gt;.&lt;/p&gt;
&lt;p&gt;Even files without secrets leak useful information. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.upguard.com/blog/yolo-mode-hidden-risks-in-claude-code-permissions&quot;&gt;UpGuard analyzed&lt;/a&gt; 18,470 Claude Code local settings files from GitHub and found that 21.3% allowed the agent to run curl freely, which an attacker who hijacks the agent through prompt injection could use to download malware or send local data out.&lt;/p&gt;
&lt;p&gt;The broader trend is moving the wrong way. GitGuardian counted 28.65 million new hardcoded secrets in public GitHub commits in 2025, up 34% from the year before and the largest annual jump it has recorded. Leaked secrets tied to AI services rose 81%, and it found 24,008 secrets inside MCP configuration files, the files that connect AI agents to outside tools, &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.turbogeek.co.uk/?p=9468&quot;&gt;according to TurboGeek’s summary&lt;/a&gt;. Old leaks also linger. The same summary reports that 64% of secrets GitGuardian first detected in 2022 were still active and unrevoked in 2025, which tells you how rarely people rotate a key after it leaks.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-06.1QHSdo63_Z1Aj3FE.webp&quot; alt=&quot;New hardcoded secrets found in public GitHub commits each year.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1488&quot; height=&quot;838&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;New hardcoded secrets found in public GitHub commits each year. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.gitguardian.com/state-of-secrets-sprawl-report-2026&quot;&gt;GitGuardian State of Secrets Sprawl 2026&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;what-the-research-leaves-out&quot;&gt;What the research leaves out&lt;/h2&gt;
&lt;p&gt;The research has limits, and they matter. Apiiro did not test whether the 2,539 secrets still work, and it has not said whether anyone has used them. Only 1 key is confirmed to have been live. The 1.1% figure counts public files, so it says nothing about how many Claude Code users have secrets in private repositories or on their laptops. Some secrets will be old test credentials that were rotated long ago.&lt;/p&gt;
&lt;p&gt;The risk is still real. Apiiro noted that in September 2025, the Shai-Hulud worm ran TruffleHog over infected developer machines, and that scanner flagged more than half of these settings files as containing secrets. Attackers already have tools that look in developer folders. Now that this research is public, more of them will look in .claude.&lt;/p&gt;
&lt;h2 id=&quot;the-lesson-for-anyone-using-an-ai-coding-agent&quot;&gt;The lesson for anyone using an AI coding agent&lt;/h2&gt;
&lt;p&gt;2 things from this generalize past Claude Code. The first is that agent folders hold secrets, and they deserve the same treatment as .env files. That covers .claude, .cursor, .gemini, .specstory and agent memory files, on laptops as well as in repositories. Add them to the project’s own .gitignore and to your organization’s gitignore templates, so the protection travels with the code to every machine that clones it.&lt;/p&gt;
&lt;p&gt;The second is that a convenience click can create a permanent record. If you use Claude Code with any public repository, search your GitHub organization today for .claude/settings.local.json and .claude/settings.json, including git history. Rotate every password, token and key those files have ever held before you clean anything up. Then stop tracking the local file in git, add it to the project’s .gitignore and open the copy on your own machine to delete any rule that contains a secret. Deal first with any credential that sits next to a public hostname, as Apiiro recommends. From then on, keep secrets out of commands entirely by reading them from a secrets manager through environment variables, approve broad command patterns such as running npm scripts so no credential ever lands in a rule, and set up commit hooks and CI secret scanning to read git history and catch passwords inside commands, starting with the 215 shapes that gitleaks and TruffleHog both missed. Claude Code’s own &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://code.claude.com/docs/en/settings&quot;&gt;settings documentation&lt;/a&gt; shows deny rules that block the agent from reading .env files and a secrets folder, and organizations can push managed settings that individual users cannot override, which is the place to enforce those rules across a team.&lt;/p&gt;</content:encoded><dc:creator>Shivashish Yadav</dc:creator><category>Security</category><category>claude</category><category>ai</category><category>ai-agent</category><category>artificial-intelligence</category><category>llm</category><author>hello@soursen.org (Shivashish Yadav)</author></item><item><title>OpenAI’s Agents Built 948 Working Linux Root Exploits in a Single Run.</title><link>https://soursen.org/articles/openais-agents-built-948-working-linux-root-exploits-in-a-single-run/</link><guid isPermaLink="true">https://soursen.org/articles/openais-agents-built-948-working-linux-root-exploits-in-a-single-run/</guid><description>An AI pipeline found, proved and drafted fixes for 948 kernel bugs at about $1,279 each. Finding them was the cheap part. The bigger lesson is who now has to fix them, and how fast.</description><pubDate>Fri, 06 Nov 2026 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;On November 4, 2026, OpenAI published a research post called &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://openai.com/index/a-thousand-fewer-paths-to-root/&quot;&gt;A thousand fewer paths to root&lt;/a&gt;, and the numbers are hard to shrug off. In a single run in mid October, a pipeline of AI agents scanned the Linux kernel, reproduced 1,724 real vulnerabilities, built 1,169 working root exploits and traced them back to 948 distinct bugs, each with a candidate patch. The whole run cost $1,213,430 in API credits, roughly $1,279 per bug. I want to walk through what OpenAI built, why the cost matters, what went wrong inside the pipeline and what the report leaves out, because the part that should worry Linux users is the gap between finding these bugs and getting fixes onto real machines. I will keep this at the level OpenAI itself published. The company withheld exploit code and technique details, and so will I.&lt;/p&gt;
&lt;h2 id=&quot;what-a-root-bug-gives-an-attacker&quot;&gt;What a root bug gives an attacker&lt;/h2&gt;
&lt;p&gt;On Linux, root is the account that can do anything. It can read every file, install software, change system settings, load kernel modules, hide processes and switch off logging. Normal user accounts, service accounts and container processes are deliberately fenced off from all of that.&lt;/p&gt;
&lt;p&gt;A local privilege escalation bug lets someone who is already running code on a machine jump that fence. When the flaw sits in the kernel, the attacker usually ends up with the highest level of control the system has. Every one of OpenAI’s 948 bugs is local, so an attacker needs some foothold first, such as a user account, an SSH login, a compromised app or a process inside a container. From there, any of them could hand over root.&lt;/p&gt;
&lt;p&gt;These bugs are old news for Linux. Dirty Pipe (&lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://recordedfuture.com/vulnerability-spotlight-dirty-pipe&quot;&gt;CVE-2022–0847&lt;/a&gt;) in 2022 let unprivileged users overwrite files they could only read, and it became the standard example of how a single kernel mistake can turn a regular login into root. What changed in 2026 is the volume and the speed.&lt;/p&gt;
&lt;h2 id=&quot;what-openai-built&quot;&gt;What OpenAI built&lt;/h2&gt;
&lt;p&gt;OpenAI built a prototype pipeline of AI agents that hunts for security bugs in the Linux kernel, checks whether they are real, builds a working exploit to prove the impact and writes a candidate patch. A human security researcher reviews the root cause analysis, the exploit and the proposed fix before anything goes to kernel maintainers.&lt;/p&gt;
&lt;p&gt;According to OpenAI, the October run used an upcoming cyber model that the company has not named or released. The agents scanned the latest Linux kernel master branch at the time. The exploits were tested against a commonly used Linux distribution, which OpenAI also did not name, running in virtual machines with a stock desktop setup and the standard kernel protections switched on, including SMEP, SMAP, KASLR and HARDENED_USERCOPY.&lt;/p&gt;
&lt;p&gt;The work is part of OpenAI’s &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://openai.com/index/daybreak-securing-the-world/&quot;&gt;Daybreak security program&lt;/a&gt;, which launched in May 2026, and its Patch the Planet initiative, added on June 22 to turn findings into merged fixes for open source maintainers.&lt;/p&gt;
&lt;h2 id=&quot;how-40-million-lines-became-948-bugs&quot;&gt;How 40 million lines became 948 bugs&lt;/h2&gt;
&lt;p&gt;The Linux kernel has more than 40 million lines of code. OpenAI pointed its agents at a slice of it, and the pipeline narrows that slice in 5 stages.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-01.oyXpyJBl_ZN9JsW.webp&quot; alt=&quot;Each stage of the October run narrows the pile, from 5,034 audit targets down to 948 distinct bugs.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1600&quot; height=&quot;902&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Each stage of the October run narrows the pile, from 5,034 audit targets down to 948 distinct bugs. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://openai.com/index/a-thousand-fewer-paths-to-root/&quot;&gt;OpenAI, Nov 2026&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;p&gt;First, Scout agents read the code and picked 5,034 audit targets. Those covered about 2.8 million lines, roughly 7% of the kernel, chosen because a low privileged user can reach them and because they touch security. That step cost $3,426.&lt;/p&gt;
&lt;p&gt;Second, Analyst agents reviewed those targets and produced 2,308 candidate findings in 12 hours, for $316,549.&lt;/p&gt;
&lt;p&gt;Third, triage agents tried to reproduce each candidate in an isolated virtual machine running a kernel built with KASAN, the kernel’s memory error detector. They reproduced 1,724 of them within 30 hours, a hit rate of about 74.7%. OpenAI kept only findings that a low privileged user could reach, that needed no special capabilities and that produced measurable evidence of memory corruption.&lt;/p&gt;
&lt;p&gt;The other 584 mostly fell away for practical reasons. OpenAI says most either required elevated privileges to begin with, or failed to reproduce inside the time budget because of difficult race conditions, very specific trigger conditions (its example is an allocation failure at a single specific line) or slow validation such as waiting for a reference counter to wrap around. Only 3.1% of that remainder reflected incorrect analysis by the model.&lt;/p&gt;
&lt;p&gt;Fourth, a screening agent sorted the reproduced bugs into 313 information leak candidates and 1,238 privilege escalation candidates. Exploit development agents then attacked them inside isolated VMs and produced 291 working leak proofs of concept and 1,169 working root exploits.&lt;/p&gt;
&lt;p&gt;Fifth, OpenAI ran variant analysis, grouping exploits that shared a root cause and a fix. That collapsed 1,169 exploits into 948 distinct vulnerabilities with 948 distinct patches. So the 948 counts bugs by the patch that closes each one, and it is OpenAI’s own figure.&lt;/p&gt;
&lt;h2 id=&quot;where-the-12-million-went&quot;&gt;Where the $1.2 million went&lt;/h2&gt;
&lt;p&gt;The exploitation stage ate most of the budget. Building working root exploits alone cost $691,000, and the information leak exploits cost another $84,929.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-02.1zM2oVzH_ZsD8qb.webp&quot; alt=&quot;API spend by stage. Building the root exploits cost more than every other stage combined.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1600&quot; height=&quot;902&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;API spend by stage. Building the root exploits cost more than every other stage combined. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://explainx.ai/blog/openai-linux-kernel-948-local-privilege-escalation-exploits-1-2-million-api-cost-pipeline-patch-the-planet-2026&quot;&gt;OpenAI stage costs via explainx.ai, Nov 2026&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;p&gt;I think this ratio tells you where the difficulty lives. Finding 2,308 candidate bugs cost about $317,000. Proving which ones actually give you root cost more than twice that. Discovery got cheap first. Proof is still the expensive step, even for machines.&lt;/p&gt;
&lt;p&gt;OpenAI also put the spend next to bug bounty prices. Competitions such as Pwn2Own and TyphoonPWN have historically offered $20,000 and $70,000 for a single Linux kernel privilege escalation exploit. Multiply those by 948 and you get between $18.9 million and $66.3 million.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-03.1K_82wpZ_Z250Prt.webp&quot; alt=&quot;OpenAI’s API spend next to what 948 exploits would cost at historical competition bounty rates.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1600&quot; height=&quot;902&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;OpenAI’s API spend next to what 948 exploits would cost at historical competition bounty rates. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.csoonline.com/article/4245265/openai-built-an-ai-pipeline-to-find-exploit-and-patch-linux-flaws.html&quot;&gt;CSO Online, Decrypt, Nov 2026&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;p&gt;&lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.csoonline.com/article/4245265/openai-built-an-ai-pipeline-to-find-exploit-and-patch-linux-flaws.html&quot;&gt;CSO Online&lt;/a&gt; points out that this is a hypothetical comparison, since nobody paid out those bounties. I still find it useful. Kernel exploit development has traditionally been slow, specialist work, which is exactly why competitions pay so much for each one. The pipeline produced 1,169 working exploits in a campaign that coverage of the post describes as a roughly 10 day window of automated scanning and manual analysis, for a price a mid sized security team could approve.&lt;/p&gt;
&lt;h2 id=&quot;5-times-the-bugs-in-3-months&quot;&gt;5 times the bugs in 3 months&lt;/h2&gt;
&lt;p&gt;OpenAI ran an earlier campaign against the same mainline kernel in late July 2026 using GPT-5.6 Sol. According to &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://decrypt.co/382632/openai-linux-security-flaws-exploit&quot;&gt;Decrypt&lt;/a&gt;, that run found 166 unique privilege escalation bugs for $513,690. The October run found 948, more than 5 times as many, even though 3 months of kernel fixes had landed in between. Spend rose about 2.4 times while the bug count rose about 5.7 times, so most of the jump came from the newer model. The cost per distinct bug fell from about $3,094 to about $1,279.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-04.Pb8Hm6tU_ZOERDr.webp&quot; alt=&quot;The October run found more than 5 times as many bugs as the July run, at less than half the cost per bug.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1600&quot; height=&quot;902&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;The October run found more than 5 times as many bugs as the July run, at less than half the cost per bug. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://decrypt.co/382632/openai-linux-security-flaws-exploit&quot;&gt;OpenAI via Decrypt, Nov 2026&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;p&gt;Go back further and the curve gets steeper. When OpenAI launched Patch the Planet in June, its GPT-5.5-Cyber model had produced 24 Linux privilege escalation exploits. June gave 24. July gave 166. October gave 948.&lt;/p&gt;
&lt;p&gt;GPT-5.5-Cyber itself is restricted. OpenAI only gives it to vetted defenders through its Trusted Access for Cyber program, and &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://siliconangle.com/2026/06/22/openai-expands-daybreak-patch-planet-full-gpt-5&quot;&gt;SiliconANGLE&lt;/a&gt; reported its score on CyberGym, a benchmark that tests whether an agent can reproduce known vulnerabilities, at 85.6% against 81.8% for standard GPT-5.5. The model behind the October run is newer than that and has no public name.&lt;/p&gt;
&lt;p&gt;OpenAI also tested the pipeline on Google’s kernelCTF, a deliberately hardened kernel with a reduced attack surface that pays bounties for working exploits. There, the pipeline produced 25 information leak proofs of concept and 21 root exploits in 112 hours, for about $32,900 in API credits. For scale, OpenAI noted that kernelCTF had accepted 26 confirmed zero day privilege escalation bugs in all of 2026 up to that point.&lt;/p&gt;
&lt;h2 id=&quot;10-techniques-the-experts-had-not-seen&quot;&gt;10 techniques the experts had not seen&lt;/h2&gt;
&lt;p&gt;OpenAI reviewed a sample of 166 exploits and found the agents had used at least 10 techniques that were new to its researchers and to the outside kernel exploitation experts it consulted. It published them only at category level, and I will do the same. There are 2 new elastic objects that persist in kernel memory and help an exploit shape the heap, 2 new page grooming techniques that make exploits more reliable, 5 new data only techniques that reach root without injecting code and 1 technique involving arbitrary type confusion.&lt;/p&gt;
&lt;p&gt;The data only group matters most to defenders. These attacks work by quietly changing kernel data, so protections such as SMEP and SMAP, which stop the kernel from executing or touching user memory, have less to catch. I expect kernel hardening work over the next year to respond to exactly this class.&lt;/p&gt;
&lt;h2 id=&quot;the-20-exploits-that-found-the-wrong-bug&quot;&gt;The 20 exploits that found the wrong bug&lt;/h2&gt;
&lt;p&gt;OpenAI also published a failure, and I think it is the most useful part of the post. In an early exploitation run, the agents reported 200 successful root exploits. When OpenAI’s researchers checked, 20 of those had reached root through a different, previously unknown bug than the one each agent had been assigned.&lt;/p&gt;
&lt;p&gt;That means 10% of the “successes” were the agents finding their own way in. It is a real capability, and it also means an automated pipeline can tell you it proved bug A when it actually proved bug B. OpenAI says it added independent checks to confirm each exploit used its intended vulnerability.&lt;/p&gt;
&lt;p&gt;Yash Kadakia, founder and CTO of Security Brigade, told CSO Online that buyers of AI security tools should demand reproducible exploits, accurate root cause analysis and independent validation. This anecdote is the reason. The patches needed human attention too. CSO reports that early AI generated patches sometimes added unneeded code changes, introduced performance trade offs or missed related variants. OpenAI says the pipeline also flagged problems in some human written patches before they merged, so the review works in both directions.&lt;/p&gt;
&lt;h2 id=&quot;only-61-fixes-have-landed&quot;&gt;Only 61 fixes have landed&lt;/h2&gt;
&lt;p&gt;At the time of the post, OpenAI reported 61 of the bugs patched upstream, 43 with patches in staging and 299 being reported and discussed publicly. That leaves roughly 545 still moving through the reporting pipeline.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-05.BxlL-72B_1jBRAc.webp&quot; alt=&quot;Disclosure status of the 948 bugs at publication. The grey segment is Soursen arithmetic.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1600&quot; height=&quot;669&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Disclosure status of the 948 bugs at publication. The grey segment is Soursen arithmetic. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.csoonline.com/article/4245265/openai-built-an-ai-pipeline-to-find-exploit-and-patch-linux-flaws.html&quot;&gt;OpenAI via CSO Online, Nov 2026&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;p&gt;OpenAI says it has already sent hundreds of reports and patches to maintainers, starting with the most critical issues and following the kernel community’s disclosure guidance. It is routing the rest through partners in Patch the Planet.&lt;/p&gt;
&lt;p&gt;“Patched upstream” only means the fix landed in the mainline kernel. Your distribution still has to backport it, build it and ship it, and you still have to install it and reboot. Sakshi Grover, research director at IDC, told CSO Online that the real exposure is “the length of the window between discovery and fixes reaching production.” She suggested enterprises prioritize multi tenant hosts, CI/CD runners, shared bastion servers, container platforms and developer fleets, and ask their Linux vendors for backport timelines.&lt;/p&gt;
&lt;h2 id=&quot;the-maintainers-carry-the-cost&quot;&gt;The maintainers carry the cost&lt;/h2&gt;
&lt;p&gt;Kernel maintainers have been saying all year that AI bug hunting lands on them. In his Linux 7.1-rc4 announcement in May 2026, Linus Torvalds said the kernel’s security mailing list had become “almost entirely unmanageable,” &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://technewsday.com/?p=56271&quot;&gt;according to Technewsday&lt;/a&gt;. The problem was duplicates. Researchers running the same AI tools kept filing the same bugs, and because the security list is private, nobody could see what had already been reported. Maintainers spent their time pointing people at fixes that had landed weeks earlier.&lt;/p&gt;
&lt;p&gt;At Open Source Summit North America on May 20, Torvalds talked through the same strain with Dirk Hohndel on stage. According to &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://lwn.net/Articles/1073782/&quot;&gt;LWN’s write up of the session&lt;/a&gt;, he said kernel commits had risen by about 20% over the last 2 releases, and he warned that when 1 person finds a bug with AI, “a hundred other people also found it with AI.” He also said maintainers of smaller projects are at risk of burnout, and Hohndel complained that bug reports from companies spending heavily on bug hunting arrived without patches. At the Open Source Summit in Prague in October, &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.digitaltoday.co.kr/en/view/112701/linus-torvalds-ai-is-an-entry-tool-for-programming-caution-needed-for-important-work&quot;&gt;Digital Today, citing TechRadar&lt;/a&gt;, reported Torvalds saying that about 3 quarters of the discussion at a maintainer summit focused on making AI tools less burdensome and more useful.&lt;/p&gt;
&lt;p&gt;Torvalds also rejected the idea that closing source code would help. As he told &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.zdnet.com/article/linus-torvalds-has-a-love-hate-relationship-with-ai/&quot;&gt;ZDNet&lt;/a&gt;, AI can reverse engineer closed source software too.&lt;/p&gt;
&lt;p&gt;Patch the Planet is OpenAI’s answer to that triage burden. It was built with Trail of Bits, with HackerOne and Calif helping on triage and disclosure, and OpenAI says a human security engineer reviews every finding before it reaches a maintainer. More than 30 projects signed up at launch, including cURL, Go and Python. Trail of Bits put its entire security research organization on the effort, and a 5 day initial sprint across 19 projects surfaced hundreds of issues and merged dozens of patches. That model sends maintainers a reviewed bug with a patch attached, which addresses the drive by AI reports Torvalds and Hohndel complained about.&lt;/p&gt;
&lt;p&gt;My view is that this is the right shape, and the numbers still worry me. 948 reviewed patches is a lot of review for a project whose subsystems are often run by 1 to 3 people, and OpenAI plans to extend the same work to other critical open source projects.&lt;/p&gt;
&lt;h2 id=&quot;part-of-a-bigger-surge&quot;&gt;Part of a bigger surge&lt;/h2&gt;
&lt;p&gt;Google Threat Intelligence Group counted 5,045 disclosed vulnerabilities in January 2026 and 10,740 in August 2026, &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://thehackernews.com/2026/11/openai-ai-pipeline-generates-948-linux.html&quot;&gt;as reported by The Hacker News&lt;/a&gt;. High risk disclosures went from 131 to 350 over the same months, a 167% increase. The UK’s National Cyber Security Centre has said AI cyber tools have almost certainly driven a significant increase in zero day discoveries.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-06.BUGbJ9FW_Z2abHIz.webp&quot; alt=&quot;Monthly vulnerability disclosures more than doubled between January and August 2026.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1600&quot; height=&quot;902&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Monthly vulnerability disclosures more than doubled between January and August 2026. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://thehackernews.com/2026/11/openai-ai-pipeline-generates-948-linux.html&quot;&gt;Google Threat Intelligence Group via The Hacker News, Nov 2026&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;p&gt;Linux has felt it directly. Copy Fail (CVE-2026–31431) is the clearest example. Theori’s AI driven platform, Xint, scanned the kernel’s crypto subsystem in about an hour and found a privilege escalation bug present in mainstream kernels since 2017, &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.techzine.eu/news/security/140968/linux-distributions-worldwide-targeted-by-the-copy-fail-exploit/&quot;&gt;according to Techzine&lt;/a&gt;. Theori reported it on March 23, the fix landed upstream on April 1 and the public proof of concept appeared on April 29. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://cyberscoop.com/copy-fail-linux-vulnerability-artificial-intelligence/&quot;&gt;CyberScoop reported&lt;/a&gt; that CISA added it to its Known Exploited Vulnerabilities catalog within days, and that hundreds of copycat AI generated exploit scripts surfaced after disclosure. Several more kernel root bugs followed through May.&lt;/p&gt;
&lt;p&gt;There is also a real example of why “local only” deserves attention. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://securityweek.com/openai-agents-exploited-linux-kernel-flaw-on-companys-own-systems&quot;&gt;SecurityWeek reported&lt;/a&gt; that on July 19, 2026, some of OpenAI’s own agents spotted that the kernel on their machine had a known public bug, CVE-2026–53362. They adapted an existing exploit, used it to break out of a container, gained root on the underlying worker node and moved further into the connected environment. CISA later added that CVE to its Known Exploited Vulnerabilities catalog and gave federal agencies until August 30 to patch it. That is the local threat model in practice. The attacker only needed to be running code on the box.&lt;/p&gt;
&lt;h2 id=&quot;what-the-report-leaves-out&quot;&gt;What the report leaves out&lt;/h2&gt;
&lt;p&gt;The report has limits, and a fair reading puts them on the table. OpenAI has not named the model that ran the October campaign, and it is unreleased. It has not named the Linux distribution the exploits were tested against. The 948 figure comes from OpenAI alone, no outside group has reproduced it, and it counts distinct patches written, which is a different number from fixes merged into the kernel. The $1.2 million covers API credits only. It leaves out the salaries of the researchers who reviewed every finding and the time maintainers will spend reviewing 948 patches. With exploit code and technique details withheld, outside experts cannot yet judge how reliable or practical each exploit is.&lt;/p&gt;
&lt;p&gt;I think withholding the exploits is the right call while most fixes are unmerged. It does mean several of the headline numbers rest on trust for now.&lt;/p&gt;
&lt;h2 id=&quot;the-lesson-for-anyone-running-linux&quot;&gt;The lesson for anyone running Linux&lt;/h2&gt;
&lt;p&gt;2 things from this generalize past the kernel. The first is that a low privileged account is worth far more to an attacker than it used to be. OpenAI’s own guidance says a low privileged account on a system that is behind on updates “should be considered functionally equivalent to root access for a well-resourced threat actor.” On a single user laptop where nobody else has a login, the immediate risk from these specific bugs is low. The risk climbs fast wherever untrusted code runs as an ordinary user every day, which covers shared servers, CI runners, Kubernetes nodes, hosting platforms and AI coding agents with shell access.&lt;/p&gt;
&lt;p&gt;The second is that the window between discovery and deployed fix now decides your exposure. Install kernel updates quickly and reboot into the new kernel, since a patched kernel on disk does nothing until it is running. Run untrusted code in stronger isolation such as Firecracker microVMs or gVisor, as OpenAI recommends, so a kernel bug inside the sandbox stays inside it. Restrict unprivileged user namespaces where your workloads allow it, a standard hardening step that security outlets covering the report also flagged. If you build software, rescan code you audited before whenever a stronger model ships, and use a model to attack your own patches for missed variants before you merge them. Thomas Di Giacomo, CTO of SUSE, told CSO Online that SUSE already pairs similar AI tools with layered defences aimed at privilege escalation and container escape. Ask your own Linux vendor how fast it backports upstream kernel security fixes, and get that number in writing this quarter.&lt;/p&gt;</content:encoded><dc:creator>Shivashish Yadav</dc:creator><category>Security</category><category>openai</category><category>ai</category><category>ai-agent</category><category>llm</category><category>linux</category><author>hello@soursen.org (Shivashish Yadav)</author></item><item><title>A Fake Captcha Asked One Berlin Employee to Paste a Command. It Cost the City 5.79 TB Data.</title><link>https://soursen.org/articles/a-fake-captcha-asked-one-berlin-employee-to-paste-a-command-it-cost-the/</link><guid isPermaLink="true">https://soursen.org/articles/a-fake-captcha-asked-one-berlin-employee-to-paste-a-command-it-cost-the/</guid><description>A fake captcha got one employee to paste a command into their own terminal, and that opened the door to 2 Senate departments. The attack is a warning about a social engineering trick that is spreading. The bigger lesson is what Berlin did after it noticed.</description><pubDate>Tue, 03 Nov 2026 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;In August 2026 the German capital got hit by a ransomware group called &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/Rhysida_%28hacker_group%29&quot;&gt;Rhysida&lt;/a&gt;, and the numbers are grim. Roughly 1.44 million files, about 5.79 terabytes, published on the dark web after Berlin refused to pay 30 Bitcoin, worth around €2 million. Court records, payroll data, passports, and vulnerability assessments of the city’s water supply, all dumped in the open. I want to walk through how it happened, because the entry point is a trick anyone can fall for, and the part that turned a breach into a disaster is a timing failure the city could have avoided.&lt;/p&gt;
&lt;h2 id=&quot;who-rhysida-is&quot;&gt;Who Rhysida is&lt;/h2&gt;
&lt;p&gt;Rhysida is not a new name to anyone who tracks ransomware. The group runs a ransomware as a service operation, where affiliates rent its tools and split the proceeds, and authorities believe it is based in Russia or a Russian speaking country, partly from Russian language comments in its code. It has a track record of hitting public institutions and critical services. The British Library in October 2023, the Chilean Army, Seattle-Tacoma International Airport, the Maryland Department of Transportation, and the City of Columbus, Ohio, where it posted more than 3 terabytes after a failed extortion attempt. That Columbus demand was also 30 Bitcoin, the same figure it put on Berlin, which tells you the number is a standard opening bid rather than a calculation of what Berlin could pay.&lt;/p&gt;
&lt;h2 id=&quot;how-one-pasted-command-opened-the-door&quot;&gt;How one pasted command opened the door&lt;/h2&gt;
&lt;p&gt;The way in was a fake captcha. An employee landed on a manipulated website showing what looked like a routine Cloudflare check, the kind everyone clicks through without thinking. Instead of a checkbox, the page quietly copied a command to the clipboard and walked the user through opening Windows Terminal and pasting it in, while the fake captcha kept displaying more steps to seem legitimate. The pasted command fetched the first stage of the malware in the background. The German cybersecurity agency, the BSI, calls this variant TerminalFix and attributes it to Rhysida.&lt;/p&gt;
&lt;p&gt;This family of attack has a name, ClickFix, and it is spreading because it sidesteps the defenses companies rely on. Email filters and endpoint tools are built to stop malicious attachments and downloads. Here the user runs the command themselves, by hand, so from the machine’s point of view a person simply opened a terminal and typed. There is no attachment to scan and no download to flag at the moment it matters. The whole attack is engineered to make a careful person feel like they are following normal instructions.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-01.CK57wnHM_1OVGbb.webp&quot; alt=&quot;How a fake captcha turns a routine click into a running command. Each step is designed to look like an ordinary robot check.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1600&quot; height=&quot;889&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;How a fake captcha turns a routine click into a running command. Each step is designed to look like an ordinary robot check. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://cyberdise.io/blog/berlin-hack-started-captcha-train-exactly/&quot;&gt;BSI, Cyberdise, Sept 2026&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;what-was-behind-the-door&quot;&gt;What was behind the door&lt;/h2&gt;
&lt;p&gt;That single machine sat inside the network of 2 Senate departments, Mobility, Transport, Climate Protection and the Environment, and Urban Development, Building and Housing. Once inside, the attackers found something that made the rest easy. Passwords were sitting in ordinary office files in plain text, including credentials for building management systems, a payment database, and administrator accounts. Reporting puts the number of files containing passwords in the thousands.&lt;/p&gt;
&lt;p&gt;Plaintext passwords in office documents are the kind of thing security teams warn about for years and that keep happening anyway. They turn a single compromised workstation into a skeleton key. Exactly how the attackers moved from that one machine to the full data holdings of both departments has not been explained publicly, but the stored credentials are the obvious bridge.&lt;/p&gt;
&lt;h2 id=&quot;the-7-days-that-cost-berlin-everything&quot;&gt;The 7 days that cost Berlin everything&lt;/h2&gt;
&lt;p&gt;This is the part that should have gone differently. Data started leaving the network on 7 August, and the Mobility department detected that outflow the same day. Berlin saw the breach on day 1. The 2 affected departments were not disconnected from the state network, the Berliner Landesnetz, until 14 August. That is a gap of 7 days between noticing data leaving and actually pulling the plug, and that week is when the archive walked out the door.&lt;/p&gt;
&lt;p&gt;Put that next to the usual numbers and the failure is clear. Sophos measured a median ransomware dwell time of 4 days in 2025, the time attackers sit undetected before acting. Mandiant’s 2026 figures put the median at 14 days, or 9 when the victim spots the breach themselves. Berlin did the hard part well, it detected the intrusion immediately. Then it took 7 days to contain what it had already seen, and in a data theft the containment speed is the whole game, because every day of access is more files gone.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-02.5T8xLJGe_Z25oFAw.webp&quot; alt=&quot;Berlin detected the breach on day 1 and took 7 days to disconnect. The exfiltration happened inside that gap.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1600&quot; height=&quot;889&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Berlin detected the breach on day 1 and took 7 days to disconnect. The exfiltration happened inside that gap. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.techtimes.com/articles/326186/20260901/berlins-seven-day-ransomware-isolation-gap-let-rhysida-steal-critical-infrastructure-data.htm&quot;&gt;Tech Times, Berlin Senate timeline, Sept 2026&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-03.DasWqc-Y_1x372N.webp&quot; alt=&quot;Berlin’s detect to isolate gap against industry dwell time benchmarks. Spotting the breach was never the weak point.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1600&quot; height=&quot;889&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Berlin’s detect to isolate gap against industry dwell time benchmarks. Spotting the breach was never the weak point. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.techtimes.com/articles/326186/20260901/berlins-seven-day-ransomware-isolation-gap-let-rhysida-steal-critical-infrastructure-data.htm&quot;&gt;Sophos, Mandiant, Tech Times, Sept 2026&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;the-money-that-was-saved-before-the-breach&quot;&gt;The money that was saved before the breach&lt;/h2&gt;
&lt;p&gt;There is a budget story underneath the timeline, and it is the part that should worry other governments. The urban development department’s attack detection software was licensed for only 5% of its devices, according to &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://1023jack.com/news/attack-detection-on-5-of-devices-how-rhysida-got-into-berlin-s-administration-an/&quot;&gt;reporting by rbb24&lt;/a&gt;. The computer that the attackers used as their entry point sat outside that coverage, and so did the servers the data was pulled from. The department said expanding the coverage was not possible for financial reasons.&lt;/p&gt;
&lt;p&gt;The financial reasons have numbers attached. Berlin’s state IT budget line fell from €32.2 million to €18.2 million for 2025, and €2 million was cut specifically from intrusion detection, per netzpolitik.org. A 2023 internal security report had reportedly called for expanding exactly the kind of detection that was later cut. I do not think anyone can prove the cuts caused this breach, and the reporting is careful to say so. The pattern is hard to miss, though. The tools that might have flagged the entry machine were the tools that got trimmed to save money the year before.&lt;/p&gt;
&lt;h2 id=&quot;berlin-refused-to-pay-and-what-that-meant&quot;&gt;Berlin refused to pay, and what that meant&lt;/h2&gt;
&lt;p&gt;On 28 August, Rhysida posted a “Berlin, Germany” listing on its leak site with a countdown and a demand for 30 Bitcoin. Mayor Kai Wegner said the same day that the city would not be blackmailed, and the state’s chief digital officer repeated the position days later, calling it an attack on the state itself. Berlin kept its money.&lt;/p&gt;
&lt;p&gt;I think refusing was the right call, and I also want to be honest that it came with a price paid by other people. When the deadline passed on 4 September, Rhysida published the files. The refusal denied criminals a payout and signaled that extortion does not work on the German state, which matters for every future target. It also meant real citizens’ court records, payroll details, and passport scans went public. There is no clean version of this choice. Paying funds the next attack and guarantees nothing, refusing protects the public purse and exposes the public. Berlin chose the second, which is the position most security agencies recommend, and the cost landed on residents.&lt;/p&gt;
&lt;h2 id=&quot;what-berlin-actually-lost&quot;&gt;What Berlin actually lost&lt;/h2&gt;
&lt;p&gt;The archive deserves a concrete accounting, with one caveat up front. Every figure here comes from Rhysida’s own claims on its leak site, and Berlin has not confirmed the scope while forensic review continues. Berlin’s own chief digital officer put the count at roughly 1.3 million files, a little below the attacker’s 1.44 million, which is a reminder to treat the group’s numbers as a ceiling rather than a measurement. With that caveat, the group claims personal information on 12,076 people, 16,389 email addresses, 11,963 phone numbers, and 148 bank account numbers. It claims more than 5,000 personnel files, passports and ID cards pulled from those records, court cases, disciplinary proceedings, and committee protocols. The exposure reached the top of the departments it hit. Among the leaked records were a copy of Transport Senator Ute Bonde’s own ID and her digital signature, which shows this was not limited to junior staff.&lt;/p&gt;
&lt;p&gt;The part that moves this from a privacy disaster to a security one is the infrastructure material. Rhysida claims more than 8,000 infrastructure files, including vulnerability assessments of Berlin’s drinking water supply, and documents tied to chemical, biological, radiological, and nuclear threat planning. A vulnerability assessment of a water grid is a map of where to hit it. Even unconfirmed, that category of document being loose on the dark web is a different order of problem than leaked email addresses.&lt;/p&gt;
&lt;p&gt;The human cost was immediate too. During the 14 to 23 August disconnection, housing benefit applications and payments were suspended for more than 50,000 households, because the systems that process them were offline. The breach leaked data and also stopped a core government service for over a week.&lt;/p&gt;
&lt;h2 id=&quot;the-lesson-that-applies-to-everyone&quot;&gt;The lesson that applies to everyone&lt;/h2&gt;
&lt;p&gt;2 things from this generalize past Berlin. The first is that the entry point needed no sophisticated exploit, just a convincing fake captcha and one person following what looked like normal instructions. The defense is boring and human, teaching people that no legitimate captcha ever asks you to open a terminal and paste a command, and treating that specific prompt as an immediate red flag. The second is that detection without fast response buys you almost nothing. Berlin saw the attack on day 1 and still lost everything, because 7 days passed before it cut the connection. For any organization holding data worth stealing, the drill worth running covers how fast you notice and, just as much, how fast you can disconnect once you do.&lt;/p&gt;</content:encoded><dc:creator>Shivashish Yadav</dc:creator><category>Security</category><category>berlin</category><category>hacking</category><category>cybersecurity</category><category>security</category><category>ai</category><author>hello@soursen.org (Shivashish Yadav)</author></item><item><title>Debian Patched a Critical Kernel Bug in a Day.</title><link>https://soursen.org/articles/debian-patched-a-critical-kernel-bug-in-a-day/</link><guid isPermaLink="true">https://soursen.org/articles/debian-patched-a-critical-kernel-bug-in-a-day/</guid><description>CVE-2026–98432 is a 10 year old Linux TCP flaw that an AI model turned into a remote root exploit. Volunteer led distributions shipped the fix fast. Canonical, which sells security subscriptions, finally responded on 3 November, with a patch that sits in a testing channel and needs manual steps to apply. The delay is one problem. The shape of the response is another.</description><pubDate>Tue, 03 Nov 2026 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;A developer took all of their Ubuntu servers offline at the start of November 2026, and wrote a furious post about why. The short version is that a critical Linux kernel bug had a fix available, Debian shipped it in days, and more than 2 weeks later Ubuntu still had not. I think the anger is mostly earned. I also think the full picture is more complicated than the outrage suggests, and both halves are worth your time if you run Linux servers.&lt;/p&gt;
&lt;h2 id=&quot;what-the-bug-actually-is&quot;&gt;What the bug actually is&lt;/h2&gt;
&lt;p&gt;CVE-2026–98432 is a use after free race condition in the Linux kernel’s IPv4 TCP connection setup. When a TCP connection completes, the kernel hands the pending request’s IP options to the new socket without copying them. If that socket gets torn down while another CPU is still using those options to answer a retransmitted connection request, the kernel ends up reading memory that has already been freed. IPv6 is not affected, because it copies the options instead of sharing them.&lt;/p&gt;
&lt;p&gt;The flaw was introduced in Linux 4.4, which shipped in January 2016, so it has been sitting in nearly every Linux server for about a decade. The kernel CVE team rated it &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://ubuntu.com/security/CVE-2026-98432&quot;&gt;CVSS 9.8&lt;/a&gt;, near the top of the scale, because the attack comes over the network, needs no privileges, and has low complexity. The upstream fix, written by longtime networking maintainer Eric Dumazet, landed in early October 2026, and the CVE went public on 15 October.&lt;/p&gt;
&lt;h2 id=&quot;the-ai-exploit-that-made-everyone-pay-attention&quot;&gt;The AI exploit that made everyone pay attention&lt;/h2&gt;
&lt;p&gt;What took this from a routine high severity bug to a story was &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.anthropic.com/&quot;&gt;Anthropic’s claim&lt;/a&gt;, published 27 October, that an unreleased internal research model wrote a working exploit after some initial prompting. The exploit gives an unauthenticated remote attacker root command execution on a server, running in 2 stages, first leaking a kernel address to defeat memory layout randomization, then overwriting a callback pointer to run attacker code. The reporter credited in the kernel fix is an Anthropic researcher, Xinyang Ge, so the same organization found the bug and demonstrated the exploit.&lt;/p&gt;
&lt;p&gt;One caveat the loudest posts tend to drop is the precondition. Anthropic’s demonstrated exploit targeted servers with a multipath TCP listener enabled, and the timing it reported was about 48 minutes to leak an address plus roughly 5 more minutes to reach command execution. As of the end of October, no independent researcher had publicly reproduced it, and Anthropic had not released the exploit code, the kernel version, or the test configuration. So this is a real and serious bug, demonstrated by a credible party, with the exact conditions for the public demonstration still narrower than the scariest framing implies.&lt;/p&gt;
&lt;p&gt;That precondition is also less exotic than it sounds, which is the part that cuts the other way. The Go Security Team warned on 29 October that Go 1.24 and later turn on MPTCP by default for listen sockets on Linux, so a Go service running inside a Kubernetes pod can open exactly the kind of listener the exploit needs, in the node’s kernel, without anyone choosing to enable it. A lot of modern infrastructure is written in Go. The config that makes this exploitable is closer to a default than to a rare setting someone has to switch on.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-01.DxEyr5bR_Z9OjmL.webp&quot; alt=&quot;How the TCP bug becomes root access, at a high level. The demonstrated chain needed a specific server configuration.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1600&quot; height=&quot;889&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;How the TCP bug becomes root access, at a high level. The demonstrated chain needed a specific server configuration. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://ubuntu.com/security/CVE-2026-98432&quot;&gt;Anthropic, kernel CVE record, Oct 2026&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;who-shipped-the-fix-and-who-did-not&quot;&gt;Who shipped the fix, and who did not&lt;/h2&gt;
&lt;p&gt;This is where Ubuntu looks bad. Debian, which is run largely by volunteers, shipped fixes fast. Debian 13 got its patched kernel on 16 October, the day after the CVE went public, and Debian 12 followed on 17 and 22 October. Oracle Linux shipped fixes on 16 October. Amazon Linux 2023 patched its main kernel on 24 October.&lt;/p&gt;
&lt;p&gt;Ubuntu did not. By the end of October, every supported Ubuntu release, 22.04, 24.04, 26.04, and 26.10, was listed with a fix in progress but nothing released, and the older 16.04, 18.04, and 20.04 were marked simply vulnerable. For a company that sells security subscriptions and markets Ubuntu with words like timely, enterprise grade, and build with confidence, shipping nothing for a 9.8 rated remote root bug more than 2 weeks after the fix existed upstream is a hard thing to defend. The developer who went public was right to be angry about that specific gap.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-02.qJM4LmNs_1aReoF.webp&quot; alt=&quot;Days from the CVE going public to a shipped fix. Others reached the normal updates channel. Ubuntu’s fix reached a testing channel on 3 November.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1600&quot; height=&quot;889&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Days from the CVE going public to a shipped fix. Others reached the normal updates channel. Ubuntu’s fix reached a testing channel on 3 November. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://tuxcare.com/blog/cve-2026-98432-linux-tcp-remote-root/&quot;&gt;TuxCare, Ubuntu, distribution advisories, Oct and Nov 2026&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;ubuntu-finally-responded-and-the-response-drew-its-own-criticism&quot;&gt;Ubuntu finally responded, and the response drew its own criticism&lt;/h2&gt;
&lt;p&gt;On 3 November, 19 days after the CVE went public, Canonical published a blog post explaining where things stood. The fix exists, for Ubuntu 22.04, 24.04, and 26.04, but it is in the &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://ubuntu.com/blog/fixes-available-in-proposed-pocket-for-cve-2026-98432&quot;&gt;proposed pocket&lt;/a&gt;, a testing channel that sits before the normal updates channel most people run. Only 26.10, the newest release, had the fix in the standard archive. The older releases, 16.04, 18.04, and 20.04, were still listed as fix pending.&lt;/p&gt;
&lt;p&gt;The proposed pocket is not where you want a critical security fix to live. Packages there are not installed by a normal update, so applying the fix means manually enabling a testing repository and running a specific install command before rebooting, rather than the ordinary cycle of updating and restarting. Canonical frames this as its new strategy for accelerating CVE fixes, letting users who run their own kernel acceptance tests opt in early. The company says the delay comes from testing on more than 2,000 certified servers and devices, and that security updates are the only appropriate mitigation it recommends.&lt;/p&gt;
&lt;p&gt;That explanation did not land well. The response under the announcement was blunt, and the technical objections in it are fair. A fix for a 9.8 rated remote root bug asking users to pull from a testing channel and run commands by hand is a strange place to end up, and several people pointed out that the proposed builds cover the standard kernels but not the hardware enablement, or HWE, kernels that many 24.04 servers run, so a chunk of users still had no fix path at all. Others noted that Arch Linux had shipped the patch almost 2 weeks earlier through a normal update, and that Canonical offers paid kernel livepatching while asking everyone else to either wait or hand install a test build. The sharpest version of the complaint is simple. The stable kernel team fixed this for free weeks ago, and the distribution that sells security subscriptions is the one telling people to install a test kernel and trust that it works.&lt;/p&gt;
&lt;h2 id=&quot;real-users-started-sounding-alarms&quot;&gt;Real users started sounding alarms&lt;/h2&gt;
&lt;p&gt;This stopped being an abstract patch delay when real operators began warning their own users. On 1 November, &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.anza.xyz/&quot;&gt;Anza&lt;/a&gt;, which develops the Agave validator client for Solana, told mainnet validators to act on CVE-2026–98432 and labeled the alert urgent. Solana validators need a public IPv4 address, which is the exposed surface, and Anza builds and runs its software on Ubuntu 24.04, one of the releases with no fix. Tim Garcia of the Solana Foundation put the problem in one sentence, that the flaw “is patched in the linux kernel but not yet patched on ubuntu.” AWS, for its part, told customers that turning off MPTCP reduces exposure but does not resolve the issue, and that a kernel update and reboot is the only complete fix.&lt;/p&gt;
&lt;p&gt;That is the concrete cost of the delay. A major piece of financial infrastructure, running on Ubuntu because it is the default choice, had to scramble because the distribution had not shipped a fix that had existed upstream for more than 2 weeks.&lt;/p&gt;
&lt;h2 id=&quot;the-caveats-that-complicate-the-clean-story&quot;&gt;The caveats that complicate the clean story&lt;/h2&gt;
&lt;p&gt;The clean version of this is volunteers beat the company. The real version is messier in 2 ways that matter.&lt;/p&gt;
&lt;p&gt;First, Ubuntu was not the only commercial distribution that lagged at first. Red Hat, the other big commercial Linux vendor, initially deferred the fix and sat at affected for over 2 weeks, which pointed to a broader problem with how commercial distributions work rather than a Canonical specific failing. Those distributions maintain their own heavily modified kernels and test extensively before shipping, so they move slower than distributions that track upstream closely. That slowness is sometimes the thing enterprise customers are paying for, and sometimes, as here, it leaves them exposed longer. The catch is that Red Hat caught up. It shipped fixed kernels for RHEL 8, 9, and 10 on 1 November, about 17 days after disclosure, and Ubuntu was still the last of the majors to move.&lt;/p&gt;
&lt;p&gt;Second, the severity itself is disputed. The kernel team rated the bug 9.8, and that is the number everyone quoted. Red Hat, Oracle, and SUSE independently rated it 8.1, and Amazon rated it 7.0, because they judged the attack complexity to be higher than the kernel team did. The bug is serious on any of those scores, but 9.8 and 8.1 are not the same emergency, and the vendors moving slower were also the vendors who assessed it as somewhat less critical. That does not excuse leaving it unpatched, though it does explain some of the difference in urgency.&lt;/p&gt;
&lt;h2 id=&quot;why-this-keeps-happening-now&quot;&gt;Why this keeps happening now&lt;/h2&gt;
&lt;p&gt;The deeper pattern is what makes this urgent. A decade old bug got found and weaponized because an AI model was pointed at the kernel’s networking code and told to look. That is going to keep happening, because the cost of this kind of analysis has collapsed. Bugs that sat undiscovered for 10 years are now getting surfaced in batches, and the moment a fix lands in the public kernel tree, the fix itself tells anyone watching exactly where the hole was.&lt;/p&gt;
&lt;p&gt;That compresses the safe window to almost nothing. Once CVE-2026–98432 was public with a patch, every day a distribution took to ship that patch was a day its users were exposed with the vulnerability fully documented. In that environment, the distributions that track upstream and ship quickly have a real advantage, and the ones that batch, test, and delay are making a tradeoff that looks worse every time the gap between disclosure and exploit shrinks.&lt;/p&gt;
&lt;h2 id=&quot;what-to-actually-do-about-it&quot;&gt;What to actually do about it&lt;/h2&gt;
&lt;p&gt;If you run Ubuntu servers, check your exposure directly rather than trusting any single take, including this one. Look at &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://ubuntu.com/security/CVE-2026-98432&quot;&gt;Ubuntu’s own CVE page&lt;/a&gt; for your release’s current status, run uname -r to see exactly which kernel you are on, and check whether you are on a standard kernel or an HWE kernel, because the proposed fix covers the standard builds and some HWE lines still had no fix listed. Decide deliberately whether to pull the fix from the testing channel now or wait for it to reach normal updates, and weigh that against your actual exposure. If you need patches to land within days of a critical CVE rather than weeks, this incident is a concrete data point for choosing a distribution that tracks upstream closely, whether that is Debian or something similar. Ubuntu being slow here does not make it uniquely broken. The real takeaway is that in a world where a new critical kernel bug can go from disclosure to working exploit in under an hour of machine time, patch speed, and the channel a fix ships to, are features you should evaluate before you commit your infrastructure to a vendor.&lt;/p&gt;</content:encoded><dc:creator>Shivashish Yadav</dc:creator><category>Security</category><category>ubuntu</category><category>linux</category><category>security</category><category>cybersecurity</category><author>hello@soursen.org (Shivashish Yadav)</author></item><item><title>Trend Micro Dropped 6 Unpatched Kernel Bugs on a Public List.</title><link>https://soursen.org/articles/trend-micro-dropped-6-unpatched-kernel-bugs-on-a-public-list/</link><guid isPermaLink="true">https://soursen.org/articles/trend-micro-dropped-6-unpatched-kernel-bugs-on-a-public-list/</guid><description>In October 2026 a major security firm posted 6 Linux kernel vulnerabilities, 2 of them rated 9.0, straight to the public mailing list with no fixes attached. The maintainers were furious. The real story is that responsible disclosure may be breaking down for open source, and AI is the reason.</description><pubDate>Tue, 03 Nov 2026 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;On 21 October 2026, Trend Micro’s Zero Day Initiative sent 6 Linux kernel vulnerability reports to the public kernel mailing list, with no patches attached. Two of them carried a &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.first.org/cvss/&quot;&gt;CVSS score of 9.0&lt;/a&gt;, which is about as serious as a rating gets. Within a day the kernel maintainers were openly angry, and within a week one of the bugs had a working public exploit getting root on several Linux distributions. I want to walk through what happened, because the blow up is the obvious story and the thing underneath it is the one that should worry anyone running open source infrastructure.&lt;/p&gt;
&lt;h2 id=&quot;what-responsible-disclosure-is-supposed-to-look-like&quot;&gt;What responsible disclosure is supposed to look like&lt;/h2&gt;
&lt;p&gt;The idea is simple and old. You find a security hole, you tell the people who can fix it, privately, and you give them time to ship a fix before the details go public. The gap between telling the vendor and telling the world is what keeps attackers from getting a head start. The Linux kernel has a dedicated private channel for exactly this, &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://docs.kernel.org/process/security-bugs.html&quot;&gt;security@kernel.org&lt;/a&gt;, documented and easy to find. The whole point of it is to discuss serious unfixed bugs out of public view until there is a patch.&lt;/p&gt;
&lt;p&gt;A firm that has worked with the kernel for years knows this channel exists. That context is what made the next part so strange.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-01.DByI1a9b_1ST3zp.webp&quot; alt=&quot;The documented private path for serious unfixed kernel bugs, next to the route ZDI actually took.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1600&quot; height=&quot;889&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;The documented private path for serious unfixed kernel bugs, next to the route ZDI actually took. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://docs.kernel.org/process/security-bugs.html&quot;&gt;Linux kernel documentation, Oct 2026&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;what-trend-micro-actually-did&quot;&gt;What Trend Micro actually did&lt;/h2&gt;
&lt;p&gt;Instead of the private list, ZDI posted the reports where anyone could read them. The 6 were a mix of serious and severe.&lt;/p&gt;
&lt;p&gt;Two KSMBD bugs led the list, both rated 9.0. ZDI-CAN-33549 was an integer underflow in the SMB Direct path that can lead to remote code execution. ZDI-CAN-33774 was a race condition in KSMBD, also remote code execution, and it needs no authentication, which means any server running KSMBD on port 445 is exposed. The other 4 were a mix of local privilege escalation and information disclosure, ZDI-CAN-33367 in the networking stack’s skb_zerocopy handling, ZDI-CAN-34068 in the VMware vmwgfx graphics driver, ZDI-CAN-32005 a stack buffer overflow in a TCP search domain query path, and ZDI-CAN-34076 an uninitialized resource leak in the NVMe target code.&lt;/p&gt;
&lt;p&gt;Only one of the 6 had a usable fix in sight, a branch targeted at Linux 7.5 for the SMB Direct bug, which is a future release, so publishing it now put the detail in the open before the fix reached anyone. The rest went out with no patch at all.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-02.wrarkSrO_Z1K2Qfa.webp&quot; alt=&quot;The 6 reports Trend Micro ZDI posted to the public list, by severity. Only one had a fix anywhere in sight.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1600&quot; height=&quot;924&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;The 6 reports Trend Micro ZDI posted to the public list, by severity. Only one had a fix anywhere in sight. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://pwn.watch/i/linux-kernel-zdi-reports-leaked-to-lkml&quot;&gt;pwn.watch, Hardware Busters, Oct 2026&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;the-maintainers-did-not-take-it-quietly&quot;&gt;The maintainers did not take it quietly&lt;/h2&gt;
&lt;p&gt;Greg Kroah-Hartman, who runs the stable kernel and is effectively second in command of the project, replied to the reports directly and without much patience. “For some reason you sent this to public lists, and did not include a patch,” he wrote, and told them plainly, “Please fix your submission process.” He said versions of that more than once across the thread.&lt;/p&gt;
&lt;p&gt;Peter Zijlstra, a longtime scheduler maintainer, was blunter. He wrote on the list that the ZDI reporters are clue resistant, that the kernel team had pointed them at the right file a gazillion times, and that they ignore the feedback every time. It was rude. It was also a reaction to a firm that had, by the maintainers’ account, been told repeatedly where security bugs go and kept sending them to the wrong place anyway.&lt;/p&gt;
&lt;p&gt;The reasonable move at that point is to apologize and fix the process. That is not what happened. ZDI’s replies argued that they had submitted close to a thousand cases to the kernel over 10 years formatted essentially this same way, that the formatting had never been a problem before, and that they would release findings at rejection or after 120 days, whichever came first. When someone used the word monkeys, ZDI objected to it as having racial connotation. The maintainers were not interested in relitigating the insult, because the formatting was never the complaint. The complaint was that serious unfixed bugs went public with no patch.&lt;/p&gt;
&lt;h2 id=&quot;the-part-that-makes-this-more-than-a-spat&quot;&gt;The part that makes this more than a spat&lt;/h2&gt;
&lt;p&gt;Here is where it stops being about one firm behaving badly. One of the 6 bugs, the skb_zerocopy privilege escalation, had already been reported to the kernel a week earlier by a completely separate group, and a fix was on the way before ZDI ever posted. Two independent parties found the same flaw within days of each other.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-03.BjJ9ZxP2_Zqq2I7.webp&quot; alt=&quot;The two shifts AI brought to kernel security, and why a private disclosure window is getting harder to hold.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1600&quot; height=&quot;889&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;The two shifts AI brought to kernel security, and why a private disclosure window is getting harder to hold. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://docs.kernel.org/process/security-bugs.html&quot;&gt;Kernel maintainer discussion, Oct 2026&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;p&gt;That used to be rare. Finding a serious kernel bug took deep expertise and a lot of time, so the odds of two people stumbling onto the exact same one in the same week were low. AI changed that math. Security analysis with AI models is now common, and if two researchers point similar models at the same code with similar prompts, there is no reason they will not surface the same bug. The barrier to finding these problems used to be skill. Now it is mostly money and compute, both of which you can rent from any large provider by the hour.&lt;/p&gt;
&lt;h2 id=&quot;why-private-disclosure-may-not-survive-open-source&quot;&gt;Why private disclosure may not survive open source&lt;/h2&gt;
&lt;p&gt;Stack those two facts together and you get a real problem for the kernel specifically. In closed source software, a vendor can quietly prepare a fix and ship it in a routine update, and the disclosure stays private until they are ready. The kernel cannot do that, because every change to it is public the moment it lands. You can word a commit vaguely, bury the security fix inside an unrelated change, or describe it as a cleanup, but the code itself is right there for anyone to read.&lt;/p&gt;
&lt;p&gt;That might have been enough to hide a fix from a casual observer once. It is not enough now, and Greg Kroah-Hartman has talked about this directly. When a security related patch lands in a public tree, that patch is effectively the disclosure. Point enough AI agents at the stream of kernel changes, looking for anything with security implications, and the sensitive fixes get spotted. Not every time, but often enough that a quiet fix stops being quiet. For a non critical bug that is a minor worry. For a critical one, it means the window between the fix appearing and attackers understanding it is shrinking toward zero, which puts enormous pressure on getting patches shipped to distributions fast.&lt;/p&gt;
&lt;h2 id=&quot;what-i-take-away-from-this&quot;&gt;What I take away from this&lt;/h2&gt;
&lt;p&gt;ZDI was wrong, plainly. You do not dump 9.0 rated remote code execution bugs on a public list with no patch, whatever your formatting history, and an experienced firm does not get to play confused about where security reports go. The maintainers were right to be angry, even if the monkeys line was a cheap shot.&lt;/p&gt;
&lt;p&gt;The uncomfortable part is that fixing ZDI’s behavior does not fix the underlying thing. Cheap, fast, AI driven bug hunting means serious flaws in open code will be found by multiple people at once, and public patches will keep broadcasting the very fixes they contain. I do not think there is a clean way to run private disclosure for a project where every change is visible and every change is being watched by machines. For anything important enough that people will aim their agents at it, the safe assumption now is that if a security bug exists, it has probably already been found, maybe more than once. That is the environment the kernel, and every large open source project, now has to defend in.&lt;/p&gt;</content:encoded><dc:creator>Shivashish Yadav</dc:creator><category>Security</category><category>kernel</category><category>trend-micro</category><category>security</category><category>cybersecurity</category><category>ai</category><author>hello@soursen.org (Shivashish Yadav)</author></item><item><title>Your AI Agent Is Probably Driving an Outdated Browser With the Sandbox Switched Off</title><link>https://soursen.org/articles/your-ai-agent-is-probably-driving-an-outdated-browser-with-the-sandbox/</link><guid isPermaLink="true">https://soursen.org/articles/your-ai-agent-is-probably-driving-an-outdated-browser-with-the-sandbox/</guid><description>Playwright bundles its own Chromium that lags Chrome stable, and it launches with the sandbox off by default. For a human running tests that is a small risk. For an AI coding agent browsing the open web on your laptop, it is a bigger one. The fix takes 2 settings.</description><pubDate>Tue, 03 Nov 2026 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;On 2 November 2026 a developer who goes by nisten posted a blunt warning for anyone running AI coding agents on their laptop with no container around them. The gist was to have the agent delete Playwright, because even the latest version bundles an outdated Chromium and runs it with the sandbox turned off. I think the delete advice is too drastic for most people, and the underlying problem he pointed at is real and worth understanding, because a lot of developers are now letting AI agents browse the live web from machines that hold their credentials.&lt;/p&gt;
&lt;h2 id=&quot;what-playwright-actually-ships&quot;&gt;What Playwright actually ships&lt;/h2&gt;
&lt;p&gt;&lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://playwright.dev/&quot;&gt;Playwright&lt;/a&gt; is Microsoft’s browser automation tool, the thing a huge share of test suites and, increasingly, AI agents use to drive a browser. It does not use whatever Chrome you have installed. It downloads and bundles its own copy of Chromium, and that copy only moves forward when Playwright itself ships a new release.&lt;/p&gt;
&lt;p&gt;That is the first half of the problem. According to nisten’s breakdown, Playwright 1.64.1 bundled Chromium 156.0.8078.4, while Chrome stable had moved to 156.0.8078.79 and .80 on 29 October. The numbers look close, but the gap is the point. Google patches Chromium security bugs and pushes them to Chrome stable fast, and your real Chrome updates itself quietly in the background. Playwright’s bundled browser does not. It stays on whatever build shipped with your Playwright version until the next Playwright release catches up. So a developer who runs the latest Playwright and assumes they are current can still be driving a browser that is weeks behind on security fixes.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-01.DRmeZwMt_ZlKEoN.webp&quot; alt=&quot;Your real Chrome updates itself. Playwright’s bundled Chromium only moves when Playwright ships a new release, so it can sit behind.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1600&quot; height=&quot;889&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Your real Chrome updates itself. Playwright’s bundled Chromium only moves when Playwright ships a new release, so it can sit behind. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://playwright.dev/&quot;&gt;nisten PSA, Chrome release data, Nov 2026&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;the-known-bugs-in-that-gap&quot;&gt;The known bugs in that gap&lt;/h2&gt;
&lt;p&gt;nisten named 4 Chromium CVEs that the bundled build was behind on, CVE-2026–118306, CVE-2026–118331, CVE-2026–118344, and CVE-2026–118348. These are real Chromium 156 vulnerabilities, with fixes that landed in a later build, 156.0.8078.73, and at least one of them, CVE-2026–118306, is a sandbox escape.&lt;/p&gt;
&lt;p&gt;I want to be careful here, because a community audit of the claim noted that whether a given Playwright install is affected depends on the exact bundled build, and that the specific tie to Playwright had not been independently confirmed at the time. So treat the precise CVE list as the reporter’s claim rather than a settled fact. The structural issue does not depend on those 4 numbers, though. A browser that updates on a slower clock than Chrome stable will always spend some stretch of time behind on patched security bugs, and that stretch is exactly when a malicious page has the advantage.&lt;/p&gt;
&lt;h2 id=&quot;the-sandbox-is-the-part-that-turns-a-bug-into-a-breach&quot;&gt;The sandbox is the part that turns a bug into a breach&lt;/h2&gt;
&lt;p&gt;The detail that makes the version gap matter is the sandbox. Playwright launches Chromium with the sandbox off by default. Its chromiumSandbox option defaults to false, so launches run as if you passed the no sandbox flag, as documented in &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://qaskills.sh/blog/playwright-browser-launch-args-hardening&quot;&gt;hardening guides&lt;/a&gt; for the tool.&lt;/p&gt;
&lt;p&gt;There is a reason for that default, and it explains why it is wrong for agents specifically. Chromium’s sandbox relies on Linux user namespaces, which do not work when the browser runs as root, and the standard way to run Playwright in continuous integration is inside a Docker container as root. Defaulting the sandbox off keeps those CI runs from breaking. The trouble is that the default follows people everywhere, including onto their laptops, where they are not root and the sandbox would work fine. A setting chosen to avoid friction in a test pipeline becomes the setting an AI agent inherits while browsing the open web on a developer’s personal machine.&lt;/p&gt;
&lt;p&gt;The sandbox is the wall between a web page’s code and the rest of your computer. A browser is built on the assumption that the page it loads is hostile, so it runs the page’s rendering in a locked down process that cannot touch your files, your other processes, or your credentials. A bug in that rendering process, on its own, stays trapped inside the sandbox. A sandbox escape is a separate, harder class of bug that breaks out of that wall. The reason modern browsers are relatively safe to point at the open web is that an attacker needs both, a rendering bug and an escape.&lt;/p&gt;
&lt;p&gt;Turn the sandbox off and you delete that second requirement. Now a single rendering bug runs with the full privileges of the browser process, which on your laptop means your user account. There is no wall left to escape, because you removed it. A community blast radius analysis of exactly this setup spelled it out, a renderer bug with the sandbox on stays in the renderer, but with the sandbox off it runs as you, with access to the sockets, terminals, and stored cookies the browser process can reach.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-02.uMA0dJPh_Z28c6xh.webp&quot; alt=&quot;With the sandbox on, a page bug stays trapped. With the sandbox off, the same bug runs as you. Playwright ships with it off.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1600&quot; height=&quot;889&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;With the sandbox on, a page bug stays trapped. With the sandbox off, the same bug runs as you. Playwright ships with it off. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://qaskills.sh/blog/playwright-browser-launch-args-hardening&quot;&gt;Community hardening analysis, Nov 2026&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;why-ai-agents-change-the-risk&quot;&gt;Why AI agents change the risk&lt;/h2&gt;
&lt;p&gt;For 10 years this was a minor worry, because of how Playwright was used. A test suite points the browser at your own staging site, content you control, so a hostile page never enters the picture. If your own test site is serving malware, an outdated browser is the least of your problems.&lt;/p&gt;
&lt;p&gt;AI coding agents break that assumption. An agent told to research something, check documentation, or reproduce a bug will browse the live, open web, on your development machine, through whatever browser tooling it has, which is very often Playwright with its defaults untouched. Now the browser is loading pages you did not write and did not vet, on a machine that holds your SSH keys, your cloud credentials, and your session cookies, with the sandbox off. The exact thing the sandbox exists to protect against, an unknown page running code on your computer, is now the normal mode of operation. One commenter put the threat model plainly, assume the browser gets popped, and do not give it anything worth stealing.&lt;/p&gt;
&lt;h2 id=&quot;this-fits-a-pattern-researchers-are-already-warning-about&quot;&gt;This fits a pattern researchers are already warning about&lt;/h2&gt;
&lt;p&gt;The Playwright default is one sharp edge, and it sits inside a larger problem that security researchers have started measuring. A University of Washington team presented a study in April 2026 that tested 7 popular agentic AI browsers, the kind that let a model act on web pages for you. 4 of the 7 could be made to bypass the same origin policy, the browser rule that stops one website from reading another’s data. The researchers ran a working proof of concept against &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://openai.com/&quot;&gt;ChatGPT Atlas&lt;/a&gt;, where one site stole information from another embedded inside it, and found conditions for similar attacks in Chrome with Gemini, Claude for Chrome, and Perplexity Comet.&lt;/p&gt;
&lt;p&gt;They named 3 failure modes that are new to agent driven browsing. The first is same origin bypass, because an agent interacts with the rule differently than a human does. The second is prompt injection, where a malicious page hides instructions in its code that tell the agent what to do, such as pulling in content from an embedded site and submitting it to a form. The third is memory poisoning, where an agent stores a malicious instruction it first refused, then acts on it later once the instruction is in its memory and its origin has blurred. One of the authors, David Kohlbrenner, put the guidance bluntly, that even savvy users should not trust these systems with access to a browser holding sensitive credentials like email or bank accounts, because they are not there yet.&lt;/p&gt;
&lt;p&gt;So the sandbox being off in Playwright is one example of a general truth. The tooling for letting AI browse the web was built before anyone was letting AI browse the web with this much access, and the safe defaults have not caught up. The Playwright fix is unusually cheap, which is the good news in this particular case.&lt;/p&gt;
&lt;h2 id=&quot;the-fix-is-2-settings-and-keeping-the-tool&quot;&gt;The fix is 2 settings, and keeping the tool&lt;/h2&gt;
&lt;p&gt;Deleting Playwright is overkill for most people, and the better answers came from the replies to the warning rather than the warning itself. There are 3 practical fixes, and they stack.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-03.WX_zPrbq_Z1rboSS.webp&quot; alt=&quot;The 3 fixes, any one of which removes the sharp edge. Deleting Playwright is not required.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1600&quot; height=&quot;889&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;The 3 fixes, any one of which removes the sharp edge. Deleting Playwright is not required. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://qaskills.sh/blog/playwright-browser-launch-args-hardening&quot;&gt;Community replies, QA Skills, Nov 2026&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;p&gt;The cleanest one is to stop using the bundled Chromium at all and point Playwright at your real, auto updating Chrome. Setting the launch channel to chrome makes Playwright use the installed, patched Chrome build instead of its own bundled copy, which closes the version gap for free. Pair that with turning the sandbox back on, by setting chromiumSandbox to true, and you have restored the 2 protections the defaults removed.&lt;/p&gt;
&lt;p&gt;The second fix is isolation. Run the agent and its browser inside a container with no real credentials mounted, so that even a full compromise reaches nothing worth taking. Several developers said this is already how they operate, agent in a container, zero real secrets, on the assumption that it will eventually get popped.&lt;/p&gt;
&lt;p&gt;The third, if you must stay on the bundled browser, is to keep it current, pinning Playwright to the newest release that carries a patched Chromium rather than trusting that latest means current. None of these is hard. The reason to do one of them is that the defaults were built for a world where a human pointed the browser at their own site, and that is not the world an AI agent operates in.&lt;/p&gt;
&lt;h2 id=&quot;what-to-actually-check-right-now&quot;&gt;What to actually check right now&lt;/h2&gt;
&lt;p&gt;If you run an AI agent that can browse, open its browser launch configuration and check 3 things. Whether it uses Playwright’s bundled Chromium or your installed Chrome, whether chromiumSandbox is set to true, and whether the whole thing runs in a container or straight on your host with your real credentials in reach. If it is bundled Chromium, sandbox off, on your host, you are running the exact configuration the warning was about. Point it at Chrome, turn the sandbox on, or box it in, and the sharp edge is gone. The agents are useful enough to keep. The defaults are just wrong for how they are now being used.&lt;/p&gt;</content:encoded><dc:creator>Shivashish Yadav</dc:creator><category>Security</category><category>ai-agent</category><category>ai</category><category>artificial-intelligence</category><category>llm</category><category>chrome</category><author>hello@soursen.org (Shivashish Yadav)</author></item><item><title>An AI Found a Critical Linux Bug, Then Wrote the Exploit in 53 Minutes</title><link>https://soursen.org/articles/an-ai-found-a-critical-linux-bug-then-wrote-the-exploit-in-53-minutes/</link><guid isPermaLink="true">https://soursen.org/articles/an-ai-found-a-critical-linux-bug-then-wrote-the-exploit-in-53-minutes/</guid><description>CVE-2026–98432 is a use after free flaw in the Linux kernel’s TCP code, rated 9.8 out of 10, that can hand a remote attacker root on a server running multipath TCP. What makes it different from the usual critical CVE is who found it. An unreleased internal AI model at Anthropic spotted the bug, then built a working root exploit in about 53 minutes. I went through what the flaw is, who is patched, and why the AI part changes the math for defenders.</description><pubDate>Sun, 01 Nov 2026 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;A new Linux kernel vulnerability, &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://nvd.nist.gov/vuln/detail/CVE-2026-98432&quot;&gt;CVE-2026–98432&lt;/a&gt;, carries a &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.first.org/cvss/&quot;&gt;CVSS severity score&lt;/a&gt; of 9.8 out of 10, which is about as bad as these get. It affects &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://ubuntu.com/security/CVE-2026-98432&quot;&gt;Ubuntu&lt;/a&gt;, Red Hat, and many other distributions, and when the news broke in late October 2026 several of them did not yet have a patch out. If you run a public facing Linux server with a particular TCP feature turned on, this is one to act on rather than file away.&lt;/p&gt;
&lt;p&gt;The flaw itself is serious, but ordinary. A race condition in kernel networking code is the kind of bug that surfaces a few times a year. The part that makes this one worth writing about is that an AI found it and then wrote the exploit, and did the exploit part in under an hour.&lt;/p&gt;
&lt;h2 id=&quot;what-the-bug-actually-is&quot;&gt;What the bug actually is&lt;/h2&gt;
&lt;p&gt;The flaw lives in the code that sets up new IPv4 TCP connections. When a connection completes its handshake, the kernel hands the IP options stored with the pending request over to the new socket without making its own copy. If that socket is torn down while another CPU core is still using those options to answer a recent connection request, the kernel ends up reading memory it has already freed. That is a &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://cwe.mitre.org/data/definitions/416.html&quot;&gt;use after free&lt;/a&gt; bug, one of the most common and most dangerous classes of memory safety flaw.&lt;/p&gt;
&lt;p&gt;The deeper cause is a race between 2 CPU cores touching the same piece of memory. One core frees it while another is still reading it. If the kernel then reuses that freed memory for something else, an attacker who can influence what lands there can steer the kernel into doing something it should never do. On a networked service, that influence can arrive as a crafted packet, which is why a memory bug in the TCP path is so much worse than the same bug somewhere quieter.&lt;/p&gt;
&lt;p&gt;IPv6 is not exposed to this, because the kernel copies the options for IPv6 connections instead of sharing them. The bug was introduced back in Linux 4.4, which means it sat in the kernel for years before anyone caught it.&lt;/p&gt;
&lt;h2 id=&quot;the-multipath-tcp-angle&quot;&gt;The multipath TCP angle&lt;/h2&gt;
&lt;p&gt;The attack that Anthropic demonstrated targeted servers running a &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/Multipath_TCP&quot;&gt;multipath TCP&lt;/a&gt; listener, usually written MPTCP. Ordinary TCP uses a single network path for a connection. MPTCP lets a connection use several paths at once, so a client could ride a wired link, Wi-Fi, and a cellular link together and hold the session steady as one drops. It is useful, and it is not on by default on most systems.&lt;/p&gt;
&lt;p&gt;One caveat matters here. Neither the CVE record nor the kernel fix actually names MPTCP. The link between the feature and this exploit comes from Anthropic’s own write up. Red Hat’s position is broader, that any exposed TCP service could let a remote attacker trigger the race, so I would not treat MPTCP as the only way in. To check whether the feature is enabled on your system, you can read the net.mptcp.enabled setting through sysctl. A value of 1 means it is on.&lt;/p&gt;
&lt;p&gt;If you are a regular desktop Linux user, you are probably not affected, because MPTCP is not something most machines turn on. The population that needs to move is server operators, especially anyone with a box facing the public internet, and anyone who once enabled MPTCP for a project and forgot.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-01.mdXaPUIR_ZrKF9h.webp&quot; alt=&quot;How the use after free happens, at a conceptual level. Two cores, one freed buffer, one unlucky overlap.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1600&quot; height=&quot;889&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;How the use after free happens, at a conceptual level. Two cores, one freed buffer, one unlucky overlap. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://nvd.nist.gov/vuln/detail/CVE-2026-98432&quot;&gt;CVE-2026–98432, kernel advisory&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;the-fix-is-2-lines-and-it-already-exists&quot;&gt;The fix is 2 lines, and it already exists&lt;/h2&gt;
&lt;p&gt;The good news is that kernel developers fixed this on 5 October 2026, about 10 days before the CVE was published. The fix is small, 2 lines in effect. It changes how the kernel frees those IP options, so that instead of releasing the memory immediately, it waits until every CPU core that might still be reading it has finished.&lt;/p&gt;
&lt;p&gt;The mechanism it uses is called RCU, read copy update, a synchronization scheme built into the kernel that lets many cores read shared data safely while guaranteeing the memory is not reclaimed out from under them. In practice the fix swaps an immediate free for a deferred one that respects any reader still in flight. The kernel security team’s advice is to update to the latest stable kernel your distribution ships rather than hand applying the 2 line change yourself, because the surrounding code matters and a partial patch can go wrong.&lt;/p&gt;
&lt;p&gt;Xinyang Ge, a researcher on Anthropic’s team, is credited in the kernel fix for reporting the bug.&lt;/p&gt;
&lt;h2 id=&quot;who-is-patched-and-who-is-not&quot;&gt;Who is patched and who is not&lt;/h2&gt;
&lt;p&gt;The fix landing upstream is only half the battle, because a distribution user gets it only when their vendor ships an updated kernel package. As of late October 2026 the picture was uneven. Debian had fixes out. Oracle Linux had a fix. Amazon Linux 2023 had a fix in its 6.18 kernel from 24 October. SUSE’s Micro and Leap builds had fixes landing.&lt;/p&gt;
&lt;p&gt;On the other side, Ubuntu had not shipped across its supported releases and marked the work in progress at high priority, with the oldest releases simply listed vulnerable. Red Hat Enterprise Linux 8, 9, and 10 were affected with no release date at first, though RHEL 7 was never affected. Several SUSE enterprise builds were still in testing. So the honest status when this broke was that the patch existed but had not reached everyone, which is the dangerous window for any disclosed flaw.&lt;/p&gt;
&lt;p&gt;The severity scores also disagreed, and the disagreement is informative rather than sloppy. The kernel CVE team rated it 9.8, treating the attack as low complexity. Red Hat initially called it moderate at 5.9, arguing a race condition like this is genuinely hard to win, so a crash is more likely than a clean takeover. Red Hat then raised its own advisory to important at 8.1 on 29 October, acknowledging the real confidentiality and integrity impact. Oracle and SUSE landed at 8.1 too, and Amazon rated it 7.0. The number you believe depends on how likely you think an attacker is to win the race reliably, and reasonable security teams disagreed on that.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-02.Vo9ImanV_ZKm5Kz.webp&quot; alt=&quot;Same bug, different scores. The kernel team and the distributions did not agree on how dangerous it is.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1600&quot; height=&quot;889&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Same bug, different scores. The kernel team and the distributions did not agree on how dangerous it is. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://thehackernews.com/2026/10/linux-kernel-tcp-flaw-with-98-cvss.html&quot;&gt;vendor advisories, Oct 2026&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;the-part-that-actually-matters-an-ai-did-this&quot;&gt;The part that actually matters, an AI did this&lt;/h2&gt;
&lt;p&gt;On 27 October 2026 Anthropic’s &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.anthropic.com/research/discovering-vulnerabilities-with-claude&quot;&gt;Frontier Red Team published a report&lt;/a&gt; describing how one of its AI models found serious vulnerabilities across real, security critical software, and in several cases built working exploits for them. Anthropic did not name the model publicly, describing it only as an unreleased, internal research model. That detail matters, because it means the capability on display here is ahead of anything you can currently use yourself. Linux was one target. The others included Google Chrome, Apple Safari, Android, and Windows.&lt;/p&gt;
&lt;p&gt;On Chrome, the model chained bugs to escape the browser sandbox and pop a calculator on the host machine, the classic proof that you have won code execution. That work alone produced 62 CVEs that were assigned and patched by Chrome 156, including 13 sandbox escapes and 21 flaws that could have compromised the renderer. On Safari, it bypassed the same origin protection to read cookies, history, and session data from any site. On Android, it pulled photos and personal files off a locked Pixel 9 over a USB connection. On Windows, it found a remotely triggerable kernel bug in the HTTP/3 code that could crash servers, tracked as &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://nvd.nist.gov/vuln/detail/CVE-2026-33096&quot;&gt;CVE-2026–33096&lt;/a&gt;. The Linux bug is the one covered here, a zero click path to remote root on an MPTCP listener.&lt;/p&gt;
&lt;p&gt;The timing is the detail I keep coming back to. For the Linux exploit, Anthropic reported that the model took about 48 minutes to leak a kernel address and defeat the memory randomization defense, then about 5 more minutes to reach command execution. Call it 53 minutes from flaw to working root exploit. A human expert can do this kind of work, but it takes days to weeks, and there are not many humans who can do it at all.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-03.BjW7uABg_dOwrb.webp&quot; alt=&quot;The exploit step, which has always been the slow and expensive part, took under an hour.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1600&quot; height=&quot;871&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;The exploit step, which has always been the slow and expensive part, took under an hour. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.anthropic.com/research/discovering-vulnerabilities-with-claude&quot;&gt;Anthropic Frontier Red Team, Oct 2026&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;why-53-minutes-changes-the-threat-model&quot;&gt;Why 53 minutes changes the threat model&lt;/h2&gt;
&lt;p&gt;Security has always run on an asymmetry of effort. Finding a deep memory corruption bug and turning it into a reliable exploit is slow, specialized, expensive work, which is the main reason most disclosed vulnerabilities never become real attacks. The exploit is harder to build than the bug is to find, and that gap is a lot of what has kept defenders alive.&lt;/p&gt;
&lt;p&gt;An AI that can do the exploit step in under an hour compresses that gap. It does not make impossible attacks possible, since a skilled human could already do this. It makes expensive attacks cheap, and it makes rare expertise common. The same capability cuts both ways. Anthropic used it to find and responsibly disclose these flaws, giving vendors 90 days to patch before any detail went public and a further 45 days after that. A different operator with the same kind of capability and no disclosure ethics is the threat this foreshadows.&lt;/p&gt;
&lt;p&gt;I do not think this means the sky is falling. The flaws found here got fixed, and defenders get to use the same tools to find bugs before attackers do. What changes is the clock. The window between a bug existing and a working exploit existing used to be the defender’s friend. It is shrinking, and the practical response is boring and correct. Patch faster, reduce what you expose to the internet, and stop treating a race condition as too hard to exploit just because it was hard last year.&lt;/p&gt;
&lt;h2 id=&quot;this-is-a-trend-not-a-one-off&quot;&gt;This is a trend, not a one off&lt;/h2&gt;
&lt;p&gt;It would be easy to read this as a single dramatic demo from one lab, and miss that it is part of a line. In November 2024 Google’s &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://thehackernews.com/2024/11/googles-ai-tool-big-sleep-finds-zero.html&quot;&gt;Big Sleep&lt;/a&gt; agent found a real zero day in SQLite, a widely used database engine, which was reported at the time as the first instance of an AI finding a previously unknown exploitable memory safety bug in real world code. Since then the results have gotten steadily more serious, and this Anthropic report, with working exploits across 5 major platforms and dozens of patched CVEs, is the current high water mark rather than an outlier.&lt;/p&gt;
&lt;p&gt;The direction is clear enough. AI systems are getting good at the two hardest parts of offensive security at once, spotting a subtle bug in a huge codebase and then proving it is exploitable. Both of those used to be gated by scarce human expertise. Neither is anymore, or at least not to the degree it was. The defensive side gets the same tools, which is the only reason this is survivable, but the people who patch slowly and the software nobody has audited in years are now exposed to a kind of scrutiny they never faced before. Some of that scrutiny will come from defenders and some from attackers, and the software does not get to choose which arrives first.&lt;/p&gt;
&lt;h2 id=&quot;what-to-actually-do&quot;&gt;What to actually do&lt;/h2&gt;
&lt;p&gt;If you run Linux servers, update to the patched kernel your distribution ships as soon as it is available, and watch your vendor’s security advisory if it is not out yet. If you cannot patch immediately and you do not need multipath TCP, disabling that feature closes the demonstrated attack path for now. Do not hand apply the 2 line kernel change on your own, because the kernel team specifically advised against it.&lt;/p&gt;
&lt;p&gt;If you run a normal desktop, check whether MPTCP is enabled, which it probably is not, and otherwise keep your system updated the way you should anyway. The bug is severe, the fix exists, and the only thing standing between the two for most people is the time it takes a vendor to ship a package and an admin to install it.&lt;/p&gt;</content:encoded><dc:creator>Shivashish Yadav</dc:creator><category>Security</category><category>linux</category><category>cybersecurity</category><category>security</category><category>ai</category><category>artificial-intelligence</category><author>hello@soursen.org (Shivashish Yadav)</author></item><item><title>A Privacy Company Called Security Caution “Loser Thinking.” The Problem Is Which Company Said It</title><link>https://soursen.org/articles/a-privacy-company-called-security-caution-loser-thinking-the-problem-is/</link><guid isPermaLink="true">https://soursen.org/articles/a-privacy-company-called-security-caution-loser-thinking-the-problem-is/</guid><description>Proton, the encrypted email company that sells itself on taking security seriously, had its CEO tell an employee that asking for a few days of security testing before shipping AI written code was “bullshit and loser thinking,” according to 404 Media. The CTO says about 90% of Proton’s code is now written by AI. I think the story matters less for what Proton did and more for the exact contradiction it exposes.</description><pubDate>Fri, 30 Oct 2026 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;Most companies pushing hard on AI are selling you convenience, and if their AI written code ships a bug, you get an annoying outage. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://proton.me/&quot;&gt;Proton&lt;/a&gt; sells you privacy. Encrypted email, a VPN, a password manager, the whole pitch is that you trust them with the things you cannot afford to have leak. So when the reporting says Proton’s own CEO called a request to security test new code before release “loser thinking,” the alarming part is which company said it. Moving fast is ordinary. A privacy company treating caution as weakness is the story.&lt;/p&gt;
&lt;p&gt;The story broke through &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.404media.co/protons-ceo-cancels-ai-re-interviews/&quot;&gt;404 Media&lt;/a&gt;, reported by &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.404media.co/author/joseph/&quot;&gt;Joseph Cox&lt;/a&gt;, built on internal Slack messages, leaked documents, audio of an all hands meeting, and 6 current employees. Proton has not disputed the quotes. I went through what was actually said, because the specifics are sharper than the headline.&lt;/p&gt;
&lt;h2 id=&quot;what-andy-yen-actually-said&quot;&gt;What Andy Yen actually said&lt;/h2&gt;
&lt;p&gt;In late August 2026, an employee posted in Proton’s security Slack channel asking the company to take a few days to run penetration testing on new code before pushing it live. Penetration testing is the practice of attacking your own software the way a hacker would, to find the holes before someone else does. For a company whose entire product is security, that is a normal, careful request.&lt;/p&gt;
&lt;p&gt;CEO &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://proton.me/author/ayen&quot;&gt;Andy Yen&lt;/a&gt; replied, according to 404 Media, that it was bullshit and loser thinking, and that of course it is possible to be fast and secure. He followed with, stop thinking always of how to be slower, and focus more on how to be faster, and separately told engineers to stop lovingly handcrafting all our code and use AI more.&lt;/p&gt;
&lt;p&gt;Read that again with the product in mind. The person asking to test before shipping was not sandbagging. They were doing the one thing a privacy company is supposed to do. Being told that caution itself is loser thinking, inside the security channel, is the part that should worry Proton’s users more than any productivity memo.&lt;/p&gt;
&lt;h2 id=&quot;the-90-number-and-what-it-does-not-tell-you&quot;&gt;The 90% number, and what it does not tell you&lt;/h2&gt;
&lt;p&gt;The other number driving the story comes from Proton’s CTO, &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://proton.me/author/bbutler&quot;&gt;Bart Butler&lt;/a&gt;. 404 Media reports that in early October Butler told staff that about 90% of Proton’s code was now written by AI, and that employees say much of it is written with Anthropic’s Claude.&lt;/p&gt;
&lt;p&gt;That figure sounds enormous, and it is close to meaningless without definition. Written by AI can mean the AI generated a first draft that a human heavily edited, or it can mean the AI wrote code that shipped with light review, and those are completely different risk profiles. Nobody has published how Proton counts it, or how much of that 90% a human actually read before it went live. The number is unaudited, and I would not put weight on it as a measure of anything except how hard the company is pushing.&lt;/p&gt;
&lt;p&gt;Anthropic, the company that makes Claude, has said similar things about its own code being mostly AI written. The difference is that Anthropic is an AI lab testing the limits of its own tools on its own product. A privacy company doing the same thing with the code that guards other people’s secrets is a different bet, with other people’s data as the stake.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-01.BBzvjbqQ_Z2uO1f4.webp&quot; alt=&quot;The claim, the caveat, and why “90% written by AI” is not the reassurance it sounds like.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1600&quot; height=&quot;889&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;The claim, the caveat, and why “90% written by AI” is not the reassurance it sounds like. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.404media.co/protons-ceo-cancels-ai-re-interviews/&quot;&gt;404 Media reporting, Oct 2026&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;the-manifesto-that-told-engineers-their-job-was-gone&quot;&gt;The manifesto that told engineers their job was gone&lt;/h2&gt;
&lt;p&gt;Around the same time, Butler posted an internal AI manifesto to Slack. The leaked version opens by telling engineers their old job is gone, and frames failure to adopt AI as an existential threat to the success of the mission itself. One line reported from it is blunt about where this goes, that Proton needs to get to the point where most code is not reviewed by humans but by AI.&lt;/p&gt;
&lt;p&gt;I actually agree with parts of the manifesto as a description of where software is heading. The argument that an engineer’s job is shifting from writing code to directing and correcting models is a real and defensible read of the moment. Plenty of serious engineering organizations are reorganizing around exactly that idea. The content is not the problem.&lt;/p&gt;
&lt;p&gt;The problem is the pairing. A manifesto saying humans should stop reviewing code, published by the same leadership that called human security review loser thinking, at a company whose product is security, reads very differently than the same words would at a company selling food delivery. Context is doing all the work here, and the context is that these are the people holding your encrypted inbox.&lt;/p&gt;
&lt;h2 id=&quot;the-re-interviews-and-why-they-landed-badly&quot;&gt;The re-interviews, and why they landed badly&lt;/h2&gt;
&lt;p&gt;The push did not stop at words. 404 Media reports that Proton planned near company wide re-interviews, where employees would have to explain how they use AI, with the metrics feeding into performance reviews and, employees believed, into whether they kept their jobs. An internal email reportedly called the interviews consequential.&lt;/p&gt;
&lt;p&gt;On the Friday the story was becoming public, Yen cancelled them. His explanation, per the reporting, was that in an unintentional way they had already served their purpose. Employees read the timing differently. One said the re-interviews were not cancelled until the investigation became public. Another called the whole response consistent with damage control.&lt;/p&gt;
&lt;p&gt;To his credit, Yen did own part of it at the all hands meeting. 404 Media quotes him saying there was a communication failure on his part, while adding that he is still going to push, and that a reorganization coming the next month would involve no loss of jobs from AI. Butler was more direct, saying we fucked up, we know we fucked up, and that using AI effectively does not mean giving up on security.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-02.rzch-_rg_24wDs4.webp&quot; alt=&quot;The sequence, from the August Slack reply to the cancelled re-interviews in late October.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1600&quot; height=&quot;853&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;The sequence, from the August Slack reply to the cancelled re-interviews in late October. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.404media.co/protons-ceo-cancels-ai-re-interviews/&quot;&gt;404 Media reporting, Oct 2026&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;the-nonprofit-that-says-it-exists-for-the-mission&quot;&gt;The nonprofit that says it exists for the mission&lt;/h2&gt;
&lt;p&gt;There is a detail that makes this sharper. In June 2024 Proton &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://techcrunch.com/2024/06/17/privacy-app-maker-proton-transitions-to-non-profit-foundation-structure/&quot;&gt;moved to a nonprofit foundation structure&lt;/a&gt;, putting the Proton Foundation in control of the company. Yen framed it at the time as insulating Proton from the pressure to maximize profit, so it could stay loyal to its users and its privacy mission ahead of any shareholder demand. The whole pitch was that the mission comes first and the money follows.&lt;/p&gt;
&lt;p&gt;Hold that next to the manifesto, which calls failing to adopt AI an existential threat to the success of the mission itself. The same word, mission, is now the reason to move faster and review less. 404 Media and others report that part of the urgency is commercial, that Proton’s revenue growth had been under pressure and the AI push is partly about protecting it. I cannot audit Proton’s internal numbers, so I will not state them as fact. What I can say is that the mission framing cuts both ways, and the version where the mission justifies shipping code with less human security review is the uncomfortable one for a company that told its users the mission was exactly what would keep them safe.&lt;/p&gt;
&lt;h2 id=&quot;why-a-privacy-company-has-less-room-than-most&quot;&gt;Why a privacy company has less room than most&lt;/h2&gt;
&lt;p&gt;Here is the thing a food delivery app can do that Proton cannot. It can ship fast, break something, apologize, and patch it, and almost nobody gets hurt. The cost of a bug is an outage.&lt;/p&gt;
&lt;p&gt;For an encrypted service, the cost of a bug is the entire product. The only reason to pay Proton instead of using free email is the promise that your messages stay private. A single security hole in the wrong place does not degrade that promise, it voids it, because privacy is binary in a way uptime is not. You are either secure or you are not, and the person whose data leaks does not care that the other 99% of the code was fine. That is why a security team asking for a few days to test is not friction to route around. At a privacy company, it is the job.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-03.Bv_D3qHw_ZBS1HT.webp&quot; alt=&quot;The same bug costs a food app an outage and costs a privacy company its whole reason to exist.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1600&quot; height=&quot;871&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;The same bug costs a food app an outage and costs a privacy company its whole reason to exist. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://proton.me/&quot;&gt;Soursen analysis&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;p&gt;This is where skipping the review step gets genuinely dangerous, because AI code has a measurable security problem. Veracode’s &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.veracode.com/press-release/ai-generated-code-poses-major-security-risks-in-nearly-half-of-all-development-tasks-veracode-research-reveals/&quot;&gt;2025 GenAI Code Security Report&lt;/a&gt; ran 80 coding tasks across more than 100 models and found that when a model could choose a secure or an insecure way to write something, it picked the insecure one 45% of the time. The flaws were not obscure. They were OWASP Top 10 issues like cross site scripting and log injection, the exact category a privacy product cannot afford. Larger models were not meaningfully safer, and security did not improve even as the models got better at writing correct looking code. The AI gets more fluent without getting more careful.&lt;/p&gt;
&lt;p&gt;None of this means AI written code is automatically insecure. It can be tested like any other code, and used well it can even find vulnerabilities faster than humans do. The issue reported at Proton is the framing. Leadership treated the testing step as the enemy. You can move fast with AI and keep the security gate. What you cannot do, at a company like this, is call the gate loser thinking, when 45% of the time the thing writing your code is reaching for the insecure option.&lt;/p&gt;
&lt;h2 id=&quot;the-quote-that-explains-the-rest&quot;&gt;The quote that explains the rest&lt;/h2&gt;
&lt;p&gt;One line from the reporting stuck with me more than the loser thinking one. 404 Media reports that after an employee raised a complaint that had surfaced on Reddit, Yen told a colleague that perception is more important than truth, the truth does not matter if you do not win on perception.&lt;/p&gt;
&lt;p&gt;For most CEOs that is a cynical but survivable thing to believe. For the CEO of a company whose only asset is being trusted with the truth of what you say in private, it is closer to a mission statement in reverse. An anonymous employee in the reporting put it more plainly, saying that whatever persona Yen has outside Proton, inside it he is a straight up bully, and another described the mood as the CEO being in full AI psychosis. Those are one sided accounts from people who are angry, and they are worth reading as that. They also do not come from nowhere.&lt;/p&gt;
&lt;h2 id=&quot;what-this-is-really-a-test-of&quot;&gt;What this is really a test of&lt;/h2&gt;
&lt;p&gt;Proton is not unusual for pushing AI hard. Almost every software company is, and the general direction, engineers directing models rather than hand writing everything, is probably correct. What makes Proton a story is that it ran the experiment out loud, in the one business where the gap between moving fast and keeping your promise is widest.&lt;/p&gt;
&lt;p&gt;The real test has nothing to do with whether Proton can get AI to write 90% of its code. Plenty of companies will clear that bar. The test is whether a company that sells trust can treat the people guarding that trust as an obstacle and still deserve the word privacy on the box. We will find out from Proton’s next security incident, or from the absence of one, and not from any manifesto.&lt;/p&gt;</content:encoded><dc:creator>Shivashish Yadav</dc:creator><category>Security</category><category>proton</category><category>ai</category><category>artificial-intelligence</category><category>claude</category><category>ai-agent</category><author>hello@soursen.org (Shivashish Yadav)</author></item><item><title>Binance Lost $6.3 Billion in About 20 Minutes. The Attack Started With One Laptop</title><link>https://soursen.org/articles/binance-lost-6-3-billion-in-about-20-minutes-the-attack-started-with/</link><guid isPermaLink="true">https://soursen.org/articles/binance-lost-6-3-billion-in-about-20-minutes-the-attack-started-with/</guid><description>On 29 October 2026 attackers drained roughly $6.3 billion from Binance hot wallets in about 20 minutes, 4 times the size of the previous record. The way in was a compromised engineer’s laptop, and the clever part was a new Ethereum feature called EIP-7702 that let them turn a hot wallet into an automatic sweeper. I went through how it happened, how much got frozen, and why this did not break crypto the way it would have a few years ago.</description><pubDate>Fri, 30 Oct 2026 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;The biggest crypto theft in history took about 20 minutes. On the evening of 29 October 2026, on chain trackers started flagging strange outflows from &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.binance.com/&quot;&gt;Binance&lt;/a&gt; hot wallets, the internet connected wallets an exchange uses to move deposits and withdrawals. Within the hour, roughly $6.3 billion was gone, and blockchain analytics firm &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.trmlabs.com/&quot;&gt;TRM Labs&lt;/a&gt; traced about $6.36 billion leaving across 10 different blockchains. For scale, the previous record, the Bybit hack in February 2025, was about $1.5 billion. This one was 4 times larger.&lt;/p&gt;
&lt;p&gt;I want to walk through the mechanics, because the how is more interesting than the number, and the number is enormous.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-01.0b9ZmKQ8_1izaqU.webp&quot; alt=&quot;Where this sits against the previous record thefts. It is about 4 times the Bybit hack.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1600&quot; height=&quot;889&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Where this sits against the previous record thefts. It is about 4 times the Bybit hack. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://airdropalert.com/blogs/binance-hack-post-mortem/&quot;&gt;public post-incident reporting, 2021 to 2026&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;they-did-not-break-down-the-front-door&quot;&gt;They did not break down the front door&lt;/h2&gt;
&lt;p&gt;Binance runs one of the best funded security operations in the industry, so the attackers did not attack the vault directly. Per Binance’s initial findings, the breach started with a compromised laptop belonging to one of its DevOps engineers, the people who maintain the company’s internal infrastructure. From that laptop, the attackers reached a service container, a piece of Binance’s back end that had the ability to submit withdrawals.&lt;/p&gt;
&lt;p&gt;Binance has since said it found additional breach points but has not detailed them, and it maintains that no wallet private keys or cold storage were touched. That distinction matters. The attackers did not steal the master keys. They got into a system that was already trusted to move money, and used it.&lt;/p&gt;
&lt;h2 id=&quot;the-clever-part-turning-a-wallet-into-a-self-emptying-machine&quot;&gt;The clever part, turning a wallet into a self emptying machine&lt;/h2&gt;
&lt;p&gt;The piece that makes this hack different is how the Ethereum funds left. The attackers used &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://eips.ethereum.org/EIPS/eip-7702&quot;&gt;EIP-7702&lt;/a&gt;, a feature that shipped with Ethereum’s Pectra upgrade on 7 May 2025 and lets an ordinary wallet temporarily borrow the behavior of a smart contract. Normally a plain wallet just holds and sends funds. EIP-7702 lets it run code for a transaction, which is genuinely useful for things like paying fees in stablecoins, batching several actions into one, or adding spending limits and recovery options.&lt;/p&gt;
&lt;p&gt;The attackers pointed that feature at a Binance hot wallet labeled Binance 14. They installed a malicious contract on it that turned the wallet into an automatic sweeper, one that drained the existing balance and forwarded any new deposits to the attacker too. According to &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.blockaid.io/&quot;&gt;Blockaid&lt;/a&gt;, the attack transaction carried 50 EIP-7702 authorizations at once, fired across 50 consecutive transaction counters to raise the odds that one would land. Only 1 needed to succeed, and 1 did.&lt;/p&gt;
&lt;p&gt;I want to be fair to the Ethereum feature here, because the easy headline is that a new upgrade got exchanges hacked. EIP-7702 has real safeguards built in. An authorization is specific to one chain by default, it can be tied to the account’s transaction counter, and the account owner can revoke it. The feature did what it was designed to do. The failure was that whoever controlled Binance 14 was in a position to sign an authorization at all, which comes back to the laptop and the service container, not to Ethereum. A powerful tool handed to the wrong signer is the whole story.&lt;/p&gt;
&lt;p&gt;At 20:24 UTC a single sweep call emptied Binance 14. Within 24 seconds, Binance 16 sent another 28,005 ETH to the same address. The attacker had effectively reprogrammed the wallets so that one operator could drain them with a click. Soon after, 780 million USDC left another hot wallet, and the Bitcoin and Tron wallets were drained too. About 18,270 BTC left a Binance Bitcoin hot wallet. Etherscan relabeled the receiving address Binance Hot Wallet Exploiter, which is about as blunt as on chain labeling gets.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-02.H9iF8y_8_Z1p2p1x.webp&quot; alt=&quot;The attack chain, from a single compromised laptop to self emptying wallets, in about 20 minutes.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1600&quot; height=&quot;853&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;The attack chain, from a single compromised laptop to self emptying wallets, in about 20 minutes. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.blockaid.io/&quot;&gt;Binance findings and Blockaid analysis, Oct 2026&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;binance-went-from-business-as-usual-to-a-full-stop&quot;&gt;Binance went from business as usual to a full stop&lt;/h2&gt;
&lt;p&gt;The public response came fast. Former CEO &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/Changpeng_Zhao&quot;&gt;Changpeng Zhao, known as CZ&lt;/a&gt;, posted that it did look like a hack, and Binance paused deposits and withdrawals globally. At 02:23 UTC on 30 October, Binance formally confirmed the incident, put the impact at about $6.3 billion, and said it was working with law enforcement. Reuters reported the exchange is cooperating with the FBI and the US Secret Service. Binance US, which runs as a separate company, said it was unaffected.&lt;/p&gt;
&lt;p&gt;Binance’s promise to users was that nobody would lose money. CEO &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/Richard_Teng&quot;&gt;Richard Teng&lt;/a&gt; and the company said all losses would be covered from its SAFU fund, the Secure Asset Fund for Users that Binance keeps for exactly this, plus the company’s own balance sheet. Binance reserved over $6.34 billion to cover it, roughly $5 billion already held in the same tokens that were stolen and the rest being bought through market and OTC channels. Co-founder Yi He said the company’s annual revenue alone is enough to absorb the hit.&lt;/p&gt;
&lt;p&gt;Run the math and the scale becomes clearer. Binance reportedly holds somewhere around $160 billion in customer assets, so a $6.3 billion loss is close to 4% of custody. Painful, and survivable for a company this size. A smaller exchange taking the same hit in percentage terms would simply be gone.&lt;/p&gt;
&lt;h2 id=&quot;the-industry-froze-half-of-it-within-hours&quot;&gt;The industry froze half of it within hours&lt;/h2&gt;
&lt;p&gt;The part that would not have happened a few years ago is what came next. Within hours, the companies that issue stablecoins started freezing the stolen funds on chain.&lt;/p&gt;
&lt;p&gt;This works because of how stablecoins are built. A stablecoin like USDT or USDC is issued by a company that controls the token’s contract, and that company can blacklist an address, which makes the tokens sitting in it stop moving entirely. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://tether.to/&quot;&gt;Tether&lt;/a&gt;, led by CEO Paolo Ardoino, froze more than 1.567 billion USDT. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.circle.com/&quot;&gt;Circle&lt;/a&gt; froze over 800 million USDC. World Liberty Financial froze 27 million of its USD1, and the Arbitrum Security Council froze about 30.5 million on Arbitrum. Binance said roughly $3 billion in total has been frozen, close to half the haul, locked up before the attacker could move it.&lt;/p&gt;
&lt;p&gt;For an industry that usually watches stolen money vanish into mixers, freezing half of a record theft within hours is a real shift. It also exposes something the crypto world does not always like to say out loud. The ability to freeze those funds means the biggest stablecoins are not censorship resistant. The same power that stopped a North Korean operation from cashing out is a power the issuers hold over everyone, every day.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-03.CANfp4w7_2jzQO.webp&quot; alt=&quot;Roughly half the stolen funds were frozen within hours. The rest, mostly ETH, is beyond any issuer’s reach.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1600&quot; height=&quot;889&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Roughly half the stolen funds were frozen within hours. The rest, mostly ETH, is beyond any issuer’s reach. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.trmlabs.com/&quot;&gt;TRM Labs, Tether, Circle, Reuters, Oct 2026&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;the-money-that-got-away&quot;&gt;The money that got away&lt;/h2&gt;
&lt;p&gt;The freeze had a hard limit. Native ETH and BTC are not issued by a company, so nobody can blacklist them. The attacker moved 195,460 ETH, worth around $450 million at the time, to a fresh wallet during the first sweep, and on chain trackers estimate the attacker still controls somewhere near 480,720 ETH, over $1 billion, that no issuer can touch. That is the portion headed for mixers and cross chain bridges, the usual path for laundering funds that cannot be frozen.&lt;/p&gt;
&lt;p&gt;So the honest scoreboard is split. Roughly $3 billion frozen, roughly $3.3 billion still in the attacker’s hands, much of it in assets that will be hard to recover. The freeze cut the attacker’s winnings by about half. It did not undo the theft.&lt;/p&gt;
&lt;h2 id=&quot;who-did-it-as-far-as-anyone-can-say&quot;&gt;Who did it, as far as anyone can say&lt;/h2&gt;
&lt;p&gt;No government or exchange has officially named the attacker. What exists is attribution from blockchain intelligence firms, which is informed analysis rather than a verdict. TRM Labs said it tied the attack with high confidence to TraderTraitor, a North Korean unit linked to the &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/Lazarus_Group&quot;&gt;Lazarus Group&lt;/a&gt;. The independent investigator ZachXBT named Lazarus as the likely attacker. Binance CEO Richard Teng has said the investigation has not identified the attackers.&lt;/p&gt;
&lt;p&gt;The case for North Korea is circumstantial but strong. The Bybit hack in February 2025, the previous record, was attributed by the FBI to North Korea, and TRM says this attack shares traits with it. Two days before this hack, Binance said it had blocked a week of North Korea linked attacks that used fake job interviews to target employees, a known Lazarus tactic. None of that is proof, and I will not state it as fact, but it is why serious firms are pointing the same direction.&lt;/p&gt;
&lt;p&gt;The reason this attribution carries weight beyond a whodunit is what the money reportedly pays for. United Nations Security Council panels have for years documented North Korea using stolen crypto to fund its nuclear and missile programs, estimating the regime pulled in around $3 billion from crypto theft over a span of years and relied on it for a large share of its foreign currency. If the same pattern holds here, the stolen billions are a state financing weapons by draining an exchange, which is exactly why the freeze response matters as more than money recovery. Every dollar locked before it reaches North Korea is a dollar that does not reach the program behind it.&lt;/p&gt;
&lt;h2 id=&quot;what-it-did-to-bitcoin&quot;&gt;What it did to Bitcoin&lt;/h2&gt;
&lt;p&gt;The hack landed on an already shaky market. For weeks Bitcoin had been holding around the $80,000 level, with buyers defending every dip. When the drain started on the Thursday evening, BTC was near $81,000. Within about an hour it was below $80,000, and once that floor broke, the leverage stacked on top of it unwound fast.&lt;/p&gt;
&lt;p&gt;&lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.coinglass.com/&quot;&gt;CoinGlass&lt;/a&gt; data shows about $1.53 billion in derivatives positions liquidated over 24 hours. About 83.8% of that was long positions, traders betting the price would rise, getting force closed as it fell. Binance alone accounted for $754.6 million of the liquidations. Bitcoin printed a low near $76,460, right around its 200 day moving average, its weakest level in weeks, before clawing back toward $78,500.&lt;/p&gt;
&lt;p&gt;So it took the single largest theft in crypto history to finally knock Bitcoin through the floor it had defended for weeks. The recovery since then is the more interesting signal. A $6.3 billion exchange hack moved the price by single digit percentages and the market kept functioning.&lt;/p&gt;
&lt;h2 id=&quot;why-this-did-not-break-crypto&quot;&gt;Why this did not break crypto&lt;/h2&gt;
&lt;p&gt;A few years ago, a $6.3 billion theft from the largest exchange would have been close to an extinction event for the industry. This time Binance absorbed it from reserves, stablecoin issuers froze half of it within hours, and Bitcoin fell a single digit percentage and started recovering the same week.&lt;/p&gt;
&lt;p&gt;That resilience is real, and it comes with a bill. The hack worked because a single engineer’s laptop could reach a system trusted to move billions, which is a centralization problem Binance will have to answer for. The recovery worked because a handful of companies can freeze money on command, which is a decentralization problem the whole industry has to sit with. The money is safe for Binance’s users this time, and the two things that saved them are the two things crypto was supposed to make unnecessary.&lt;/p&gt;</content:encoded><dc:creator>Shivashish Yadav</dc:creator><category>Security</category><category>binance</category><category>cryptocurrency</category><category>crypto</category><category>web3</category><category>ai</category><author>hello@soursen.org (Shivashish Yadav)</author></item><item><title>Microsoft Keeps Closing the Local Account Loopholes in Windows 11, and the Age API Is the Reason…</title><link>https://soursen.org/articles/microsoft-keeps-closing-the-local-account-loopholes-in-windows-11-and/</link><guid isPermaLink="true">https://soursen.org/articles/microsoft-keeps-closing-the-local-account-loopholes-in-windows-11-and/</guid><description>In October 2026 Microsoft quietly killed the last easy way to set up Windows 11 without a Microsoft account, after spending 2025 blocking the rest. I went through exactly which bypasses are gone, what still works, and why a new feature called the Windows age API, that reports whether you are a child, a teen, or an adult, is the piece that ties it all together. None of this requires a conspiracy to explain, the mechanism is on the record.</description><pubDate>Tue, 27 Oct 2026 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;For years, setting up a new Windows PC with a plain local account, no email, no cloud sign in, was a known trick that enthusiasts passed around. Type a command, click a hidden link, and you could finish setup with an account that lived only on your machine. Over 2025 and 2026 Microsoft has closed those doors one by one, and the most recent closure landed in October 2026. I want to lay out what actually changed, because the facts are more specific, and more interesting, than the outrage around them.&lt;/p&gt;
&lt;h2 id=&quot;why-this-matters-to-so-many-people-right-now&quot;&gt;Why this matters to so many people right now&lt;/h2&gt;
&lt;p&gt;The timing is not an accident. Windows 10 reached &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.theregister.com/2025/10/09/business_windows_10_eol/&quot;&gt;end of support on 14 October 2025&lt;/a&gt;, which means no more security updates for the version that, as recently as July 2025, still ran on about 45% of the world’s desktops. Windows 11 only overtook it in mid 2025, reaching roughly &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.computing.co.uk/news/2025/windows-11-finally-overtakes-windows-10&quot;&gt;52% share against Windows 10’s 45%&lt;/a&gt; in StatCounter’s July 2025 numbers. So a very large population of people is being pushed onto Windows 11 right now, either by buying a new machine or by upgrading an old one, and every one of them meets the new setup flow. The account requirement is not a niche annoyance for enthusiasts, it is the first thing a wave of ordinary upgraders runs into.&lt;/p&gt;
&lt;h2 id=&quot;what-microsoft-has-actually-removed&quot;&gt;What Microsoft has actually removed&lt;/h2&gt;
&lt;p&gt;The bypasses fell in waves. In an Insider build announced on &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://alternativeto.net/news/2025/10/windows-11-now-blocks-all-microsoft-account-bypasses-during-setup/&quot;&gt;6 October 2025&lt;/a&gt;, Microsoft said it was removing known mechanisms for creating a local account during setup. Testing of that build found that the popular command line tricks were dead, oobe\bypassnro, the start ms-cxh:localonly command, the BypassNRO registry value, and even the old habit of typing a fake email to force an error were all blocked.&lt;/p&gt;
&lt;p&gt;Then on &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.windowslatest.com/2026/10/13/one-click-let-you-escape-microsofts-forced-sign-in-on-windows-11/&quot;&gt;9 October 2026&lt;/a&gt;, a researcher who goes by Bob Pony noticed the last easy one was gone. There had been a “Learn more” link on the Microsoft account screen that, when clicked, let you skip sign in without touching a command line. On a fresh 26H2 install it no longer skips anything, it just opens a data management page whose only button sends you back to the sign in screen. Microsoft closed it server side, with no mention in any update notes.&lt;/p&gt;
&lt;p&gt;So the simple, click only escape hatches are gone. A couple of harder routes still work for now. On Windows 11 Pro you can use Domain Join to create a local account, and you can build a custom install USB with a tool like Rufus that strips the Microsoft account requirement before you ever boot. Some older public install images also still accept the old commands, because the blocks appeared in Insider builds first and spread from there. Those are unsupported workarounds, and Microsoft has shown it will close them when they get attention.&lt;/p&gt;
&lt;p&gt;There is a regional carve out worth knowing. In the European Economic Area, Microsoft has loosened several Windows requirements to comply with the &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.windowscentral.com/software-apps/windows-11/microsoft-will-let-windows-11-users-in-europe-uninstall-edge-bing-and-disable-ads-in-eea-dma&quot;&gt;Digital Markets Act&lt;/a&gt;, the EU competition law that forces large platforms to give users more control. That is why reports of a local account option surviving tend to come from EU machines. Where regulators have teeth, the option stays. Where they do not, it disappears.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-01.1WnLmVcq_1NtxP6.webp&quot; alt=&quot;Figure 1. How the local account bypasses fell. Through 2025 Microsoft blocked the command line tricks like bypassnro and ms-cxh:localonly, and in October 2026 it killed the one click “Learn more” skip. Domain Join and a custom Rufus install still work for now, and the EU keeps a carve out under the Digital Markets Act.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1600&quot; height=&quot;853&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Figure 1. How the local account bypasses fell. Through 2025 Microsoft blocked the command line tricks like bypassnro and ms-cxh:localonly, and in October 2026 it killed the one click “Learn more” skip. Domain Join and a custom Rufus install still work for now, and the EU keeps a carve out under the Digital Markets Act. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.windowslatest.com/2026/10/13/one-click-let-you-escape-microsofts-forced-sign-in-on-windows-11/&quot;&gt;Author timeline based on Windows Latest and AlternativeTo reporting&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;the-official-reason-and-the-fuller-one&quot;&gt;The official reason, and the fuller one&lt;/h2&gt;
&lt;p&gt;Microsoft’s stated reason is that a Microsoft account gives you a machine that is “fully configured,” tied into OneDrive, Microsoft 365, Outlook, Xbox, the Microsoft Store, Copilot, cloud backups, cross device syncing, and Windows activation. All of that is true. A Microsoft account does connect your PC to that ecosystem from the first boot, and for people who want those services it is genuinely more convenient.&lt;/p&gt;
&lt;p&gt;The part the convenience story leaves out is what the account does for Microsoft. A signed in account ties every machine to a verified individual, which makes your activity on that computer attributable to a real person rather than just a device. That matters for data, and in 2026 it matters for a specific new feature.&lt;/p&gt;
&lt;h2 id=&quot;the-windows-age-api-is-the-piece-that-explains-the-push&quot;&gt;The Windows age API is the piece that explains the push&lt;/h2&gt;
&lt;p&gt;In late 2026 Microsoft started adding what it calls age assurance to Windows, through a pair of developer APIs. An app can call &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.bleepingcomputer.com/news/microsoft/microsoft-adds-age-awareness-apis-that-can-tell-if-users-are-children-teens-or-adults/&quot;&gt;GetUserAgeRangeAsync&lt;/a&gt; to learn which age bracket you fall into, and GetAgeVerificationStatusAsync to learn whether that age was actually verified or just self reported. The brackets are under 10, 10 to 12, 13 to 15, 16 to 17, and 18 plus. Apps do not see your date of birth, only the range, and whether it is confirmed.&lt;/p&gt;
&lt;p&gt;That age lives in your Microsoft account. Microsoft stores the verified status with the account through a system it calls Microsoft Age Verification. So the age feature only works if you are signed into a Microsoft account, which is exactly the thing setup now pushes everyone toward. The 2 stories line up. Closing the local account loopholes and rolling out an account bound age signal are the same project seen from 2 angles.&lt;/p&gt;
&lt;p&gt;Why build this now? Governments around the world are passing online age rules, and platforms need a way to tell a child from an adult at the operating system level rather than asking every single app to figure it out. An OS level age signal is Microsoft’s answer. Whether you think that is a reasonable response to real child safety laws or an overreach depends on your politics, and reasonable people land in both camps. The mechanism itself is not in dispute.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-02.CNrH4d_z_kdEAk.webp&quot; alt=&quot;Figure 2. How the Windows age API works. An app calls GetUserAgeRangeAsync and gets back 1 of 5 brackets, under 10 through 18 plus, plus whether the age was verified. The value lives in your Microsoft account through Microsoft Age Verification, which is why setup now steers everyone onto an account.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1600&quot; height=&quot;871&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Figure 2. How the Windows age API works. An app calls GetUserAgeRangeAsync and gets back 1 of 5 brackets, under 10 through 18 plus, plus whether the age was verified. The value lives in your Microsoft account through Microsoft Age Verification, which is why setup now steers everyone onto an account. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.bleepingcomputer.com/news/microsoft/microsoft-adds-age-awareness-apis-that-can-tell-if-users-are-children-teens-or-adults/&quot;&gt;Author diagram based on Microsoft developer documentation and BleepingComputer&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;how-age-verification-actually-works-and-where-it-is-live&quot;&gt;How age verification actually works, and where it is live&lt;/h2&gt;
&lt;p&gt;Self reporting your age is one thing. Verifying it is another, and this is where the privacy tradeoffs get real. When an app or a region requires a confirmed age, Microsoft routes you to its Age Verification flow, and the methods vary by country. Reported options include a facial age estimate from a selfie, uploading a government identity document, or in some flows confirming you are over 18 with a credit card. Different regions use different third party verification partners, chosen by where you are.&lt;/p&gt;
&lt;p&gt;This is not switched on everywhere. As of the rollout, age verification through Microsoft storefronts is live in &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.pcworld.com/article/3227424/windows-11-will-soon-tell-apps-how-old-you-are-heres-how-it-works.html&quot;&gt;Singapore, Brazil, and Australia&lt;/a&gt;, with more regions planned, and the underlying age APIs are broadly available to Windows Insiders ahead of a wider release. The feature is real and shipping today, though how much of it you encounter still depends on where you live and which apps adopt it.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-03.JwfapLvz_Z10v19G.webp&quot; alt=&quot;Figure 3. How you verify an age, and where it is live. The methods reported include a facial age selfie, a government ID upload, or a credit card in some flows, with the partner chosen by region. Verification through Microsoft storefronts is live in Singapore, Brazil, and Australia so far.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1600&quot; height=&quot;818&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Figure 3. How you verify an age, and where it is live. The methods reported include a facial age selfie, a government ID upload, or a credit card in some flows, with the partner chosen by region. Verification through Microsoft storefronts is live in Singapore, Brazil, and Australia so far. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.pcworld.com/article/3227424/windows-11-will-soon-tell-apps-how-old-you-are-heres-how-it-works.html&quot;&gt;Author summary based on PCWorld and Microsoft documentation&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;p&gt;The uncomfortable math is simple. A facial scan or an uploaded ID is far more identifying than an email address ever was. If age checks become common and they run through your Microsoft account, then using Windows increasingly means being a verified, named individual to Microsoft and to whoever it is legally required to answer. That is a real shift in what a personal computer is, from a device you own to an account you log into.&lt;/p&gt;
&lt;h2 id=&quot;why-the-account-model-matters-for-your-data&quot;&gt;Why the account model matters for your data&lt;/h2&gt;
&lt;p&gt;There is a knock on effect people miss. Once your machine is tied to a Microsoft account and your files sync to OneDrive, the question of who can read those files becomes important. Microsoft does not offer end to end encryption for consumer OneDrive data, which means Microsoft holds the keys and can access the files you store there, and can hand them to law enforcement under a lawful request. Apple, by contrast, offers Advanced Data Protection that even Apple cannot read. For a cloud bound account model, that gap is the whole ballgame, the more your computing lives inside an account you do not control the encryption for, the less of it is truly private.&lt;/p&gt;
&lt;p&gt;There is a second lever that points the same way. Starting with the 24H2 release, a clean install of Windows 11 &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://tech.slashdot.org/story/24/08/14/1559240/microsoft-is-enabling-bitlocker-device-encryption-by-default-on-windows-11&quot;&gt;turns on BitLocker device encryption automatically&lt;/a&gt; the first time you sign in, and it backs up the recovery key to your Microsoft or work account. Microsoft also dropped the old hardware requirements for this, so it now reaches ordinary Windows 11 Home machines too. Encrypting the disk is good for security, a stolen laptop stays locked. The catch is where the key goes. Once the recovery key sits in your Microsoft account, Microsoft can produce it, and so can anyone who compels Microsoft to. The encryption protects you from a thief and leaves you readable to the company and to a lawful request. That tradeoff is reasonable for many people, and it is one more reason the account has become the thing your whole machine hangs off.&lt;/p&gt;
&lt;p&gt;None of this means a Microsoft account is useless or that everyone should avoid it. The sync, the backups, and the cross device features are real conveniences that many people reasonably want. The honest framing is that the account is no longer optional in the way it used to be, that it now carries an age signal tied to your identity, and that the data inside it is readable by Microsoft. Knowing that is the difference between choosing the account and being quietly funneled into it.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-04.CJDPHL8G_16t6Nm.webp&quot; alt=&quot;Figure 4. Why the Microsoft account has become the anchor. Three separate changes all run through it, setup now forces the account, the age API stores your verified age in it, and 24H2 BitLocker backs up your disk recovery key to it. Each is defensible on its own, and together they make the account the thing your whole machine depends on.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1600&quot; height=&quot;871&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Figure 4. Why the Microsoft account has become the anchor. Three separate changes all run through it, setup now forces the account, the age API stores your verified age in it, and 24H2 BitLocker backs up your disk recovery key to it. Each is defensible on its own, and together they make the account the thing your whole machine depends on. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.bleepingcomputer.com/news/microsoft/microsoft-adds-age-awareness-apis-that-can-tell-if-users-are-children-teens-or-adults/&quot;&gt;Author diagram based on Microsoft documentation and reporting&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;what-you-can-actually-do-about-it&quot;&gt;What you can actually do about it&lt;/h2&gt;
&lt;p&gt;If you want a Windows machine with a local account today, the realistic paths are narrow and may not last. On Windows 11 Pro, Domain Join still creates a local account during setup. A custom install USB built with Rufus can remove the Microsoft account requirement before setup runs, and that is the most reliable method as of now. If you are in the EU or EEA, your machine may still offer a local account option directly, thanks to the Digital Markets Act. On any machine, you can create the Microsoft account during setup and then add a local administrator account afterward and switch to it, though that still means the account existed at least once.&lt;/p&gt;
&lt;p&gt;The larger choice is the one the video that prompted this gets to, if a computer with no mandatory vendor account and no account bound age signal is what you want, that path now runs through Linux, where account creation, cloud sync, and age features are all optional. That is a bigger jump, and not the right call for everyone. For most people the practical takeaway is smaller, understand that Windows 11 setup is now built to put you on a Microsoft account, that the account is becoming the anchor for an age identity, and that the convenience and the tracking are 2 sides of the same sign in.&lt;/p&gt;</content:encoded><dc:creator>Shivashish Yadav</dc:creator><category>Security</category><category>microsoft</category><category>windows</category><category>ai</category><category>artificial-intelligence</category><author>hello@soursen.org (Shivashish Yadav)</author></item><item><title>Mac vs Windows on Privacy, What the Settings Actually Do, and Where Each One Leaks</title><link>https://soursen.org/articles/mac-vs-windows-on-privacy-what-the-settings-actually-do-and-where-each/</link><guid isPermaLink="true">https://soursen.org/articles/mac-vs-windows-on-privacy-what-the-settings-actually-do-and-where-each/</guid><description>Most people run either Windows 11 or macOS, and both collect data, hand some of it to governments, and run advertising businesses. I went through what each system actually knows about you, what you can switch off, and where the real differences are, using the published numbers rather than brand loyalty. If privacy is your top priority the honest answer is neither of them, and I will get to that.</description><pubDate>Mon, 26 Oct 2026 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;If you are reading this on a laptop, it is almost certainly running Windows or macOS. Together they cover the overwhelming majority of personal computers, so for most people the privacy question is a straight fight between the 2. Wiping the machine and installing Linux changes the game entirely, and I will come back to that at the end, but first the comparison people actually face.&lt;/p&gt;
&lt;p&gt;I want to go through this the way you would evaluate it yourself, what the system collects, who it shares with, how it protects your data at rest, and how much work it takes to lock each one down. Both companies have real privacy features and both have been caught doing things their users did not expect. This piece sticks to what each one actually does with your data.&lt;/p&gt;
&lt;h2 id=&quot;what-each-system-knows-about-you&quot;&gt;What each system knows about you&lt;/h2&gt;
&lt;p&gt;Windows sends telemetry back to Microsoft in 2 tiers. The first is what Microsoft calls required diagnostic data, the basic signals about whether your machine is working. The second is optional diagnostic data, which can reach much further, the apps you open, the devices you plug in, and aspects of your browsing. You can reduce the optional tier, but understanding exactly what is being sent, and turning it down, takes real effort because the controls are spread across several settings pages.&lt;/p&gt;
&lt;p&gt;The feature that drew the most alarm is &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.bbc.com/news/articles/cx00xov9d3jo&quot;&gt;Recall&lt;/a&gt;, announced in 2024 for Copilot Plus PCs. Recall takes near constant snapshots of your screen so you can search your own past activity later. The original plan had it enabled by default, and the backlash was immediate, because a computer quietly screenshotting everything will also screenshot your banking pages, private messages, and anything a family member does on a shared machine, including people who never agreed to it. After security researchers showed the snapshots were poorly protected, Microsoft &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.geekwire.com/2024/microsoft-updates-recall-feature-after-security-and-privacy-backlash/&quot;&gt;reworked Recall to be opt in&lt;/a&gt;, encrypted the data, and tied access to Windows Hello authentication. It is off unless you choose it now, which is a real improvement, though the fact that the first design shipped on by default tells you something about the defaults mindset.&lt;/p&gt;
&lt;p&gt;Apple collects diagnostic data and user information too, and the setup process lets you decline analytics sharing, Siri, and Apple Intelligence as you go. Apple has also been caught. In 2019 a &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.theguardian.com/technology/2019/jul/26/apple-contractors-regularly-hear-confidential-details-on-siri-recordings&quot;&gt;whistleblower revealed&lt;/a&gt; that contractors grading Siri performance were listening to recordings that included medical conversations, what sounded like drug deals, and people having sex, much of it captured by accidental Siri activations. Apple apologized, said it had not been living up to its ideals, paused the grading program, and later made it opt in. The episode ended in a &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.reuters.com/legal/apple-pay-95-million-settle-lawsuit-accusing-siri-snooping-2025-01-02/&quot;&gt;$95 million settlement&lt;/a&gt; in early 2025 over the Siri recordings, paying affected users up to $20 per device, without Apple admitting wrongdoing.&lt;/p&gt;
&lt;p&gt;So both have slipped. The practical difference is how easy each makes it to say no. On a Mac the privacy toggles sit together and are clearly labeled. On Windows they are scattered, and some, like the old Recall default, started in the wrong position.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-01.GeUbZe1d_ZObJQX.webp&quot; alt=&quot;Figure 1. What each system collects and how you limit it. Windows splits telemetry into required and optional tiers and introduced Recall screen snapshots, now opt in. macOS lets you decline analytics, Siri and Apple Intelligence at setup, and settled a Siri recordings case for $95 million in 2025.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1107&quot; height=&quot;627&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Figure 1. What each system collects and how you limit it. Windows splits telemetry into required and optional tiers and introduced Recall screen snapshots, now opt in. macOS lets you decline analytics, Siri and Apple Intelligence at setup, and settled a Siri recordings case for $95 million in 2025. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.bbc.com/news/articles/cx00xov9d3jo&quot;&gt;Author summary based on Microsoft and Apple documentation and reporting&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;who-gets-your-data-when-the-government-asks&quot;&gt;Who gets your data when the government asks&lt;/h2&gt;
&lt;p&gt;This is the part people skip, and it matters more than most settings. Both companies hand user data to law enforcement, and both publish transparency reports, so you can check the numbers rather than guess.&lt;/p&gt;
&lt;p&gt;Microsoft receives tens of thousands of law enforcement requests for consumer data every year and &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.microsoft.com/en-us/corporate-responsibility/law-enforcement-requests-report&quot;&gt;publishes the totals&lt;/a&gt; in its Law Enforcement Requests Report. Apple does the same in its own &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.apple.com/legal/transparency/&quot;&gt;transparency report&lt;/a&gt;. For the first half of 2024, Apple reported 12,043 device requests from US authorities covering 42,747 devices, and said it provided data in about 85% of those US cases. The exact figures move every reporting period. What matters is that if a government asks either company for your data through the proper legal channel, you should assume it can often be produced.&lt;/p&gt;
&lt;p&gt;Where the 2 diverge is the appetite for your data in the first place. Microsoft is a sprawling ecosystem, Windows, Microsoft 365, OneDrive, Xbox, LinkedIn, Copilot, and it is actively hungry for data to feed its AI products. LinkedIn, which Microsoft owns, defaulted users into letting their posts and activity train AI. Microsoft has also been steadily tightening Windows 11 setup so that creating a plain local account is harder, pushing you toward a Microsoft account and an internet connection during install, with the old bypass commands removed in recent builds. A Microsoft account ties your machine to the company’s cloud from the first boot and makes it easier to nudge you into OneDrive and Microsoft 365. Apple runs an advertising business too, but it does not track you across unrelated apps or sell your profile to third parties. Apple’s ad system places you into audience segments and will only serve a personalized ad if &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.apple.com/legal/privacy/data/en/apple-advertising/&quot;&gt;more than 5,000 people&lt;/a&gt; match the same criteria, so you are hidden in a crowd rather than singled out. Most of the wider ad industry targets all the way down to the individual. That design choice is the clearest privacy gap between the 2 platforms, and it favors Apple.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-02.5dQnBZeK_Zez5fw.webp&quot; alt=&quot;Figure 2. Both companies disclose law enforcement requests, Apple reported 12,043 US device requests in the first half of 2024 with about 85% provided. The sharper difference is the business model, Microsoft collects broadly to feed AI products, while Apple limits ad targeting to groups of more than 5,000 people and does not track across apps.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1181&quot; height=&quot;632&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Figure 2. Both companies disclose law enforcement requests, Apple reported 12,043 US device requests in the first half of 2024 with about 85% provided. The sharper difference is the business model, Microsoft collects broadly to feed AI products, while Apple limits ad targeting to groups of more than 5,000 people and does not track across apps. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.apple.com/legal/transparency/&quot;&gt;Author chart based on Apple and Microsoft transparency reports&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;encryption-and-the-keys-that-protect-it&quot;&gt;Encryption, and the keys that protect it&lt;/h2&gt;
&lt;p&gt;Both systems can encrypt your disk, so a lost or stolen laptop does not hand over your files. On a Mac the feature is FileVault. On Windows it is BitLocker, or device encryption on some editions.&lt;/p&gt;
&lt;p&gt;The difference is in defaults and key handling. FileVault, once enabled, encrypts the whole Mac, and Apple pushes you to turn it on. BitLocker gives you the option to store your recovery key in Microsoft’s cloud, which is convenient if you forget it, and also means the key now sits on Microsoft servers, a target for anyone who breaks in there. Whether that tradeoff is acceptable depends on your threat model, but you should know the key left your machine. One important limit applies to both, disk encryption protects a powered off device. Once the machine is on and unlocked, the data is decrypted and exposed to whatever is running.&lt;/p&gt;
&lt;p&gt;Apple adds one more layer that Windows has no direct equivalent for. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.bbc.com/news/technology-62059606&quot;&gt;Lockdown Mode&lt;/a&gt;, introduced in 2022, is a stripped down high security state for people who might be targeted by mercenary spyware like Pegasus, journalists, activists, political figures. It disables risky features that attackers exploit, and Apple built it into iPhone, iPad, and Mac. It is overkill for most people and genuinely useful for the few who need it, and nothing on the Windows side matches it.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-03.B_faEyKc_ZxqkjS.webp&quot; alt=&quot;Figure 3. Disk encryption on both, with different defaults. FileVault encrypts the whole Mac and Apple pushes you to enable it, while BitLocker can escrow your recovery key to Microsoft’s cloud. Apple adds Lockdown Mode, a high security state against mercenary spyware, with no direct Windows equivalent. All disk encryption protects only a powered off device.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1094&quot; height=&quot;601&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Figure 3. Disk encryption on both, with different defaults. FileVault encrypts the whole Mac and Apple pushes you to enable it, while BitLocker can escrow your recovery key to Microsoft’s cloud. Apple adds Lockdown Mode, a high security state against mercenary spyware, with no direct Windows equivalent. All disk encryption protects only a powered off device. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.bbc.com/news/technology-62059606&quot;&gt;Author summary of FileVault, BitLocker and Lockdown Mode&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;the-icloud-encryption-fight-that-shows-the-real-limits&quot;&gt;The iCloud encryption fight that shows the real limits&lt;/h2&gt;
&lt;p&gt;There is one more piece that cuts through the marketing on both sides, and it happened to Apple in 2025. Apple offers an optional iCloud setting called &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.bleepingcomputer.com/news/security/apple-pulls-icloud-end-to-end-encryption-feature-in-the-uk&quot;&gt;Advanced Data Protection&lt;/a&gt;, introduced in December 2022, that applies end to end encryption to most of your iCloud data, backups, photos, notes, and more, so that only you, on your trusted devices, can decrypt it. With it on, Apple itself cannot read that data, which also means Apple cannot hand it to a government even if ordered to.&lt;/p&gt;
&lt;p&gt;In early 2025 the UK government served Apple with a secret order, reported by the Washington Post, demanding a backdoor that would give access to the unencrypted data of Apple users worldwide. Apple’s response was to pull Advanced Data Protection for UK users rather than build the backdoor, stopping new UK sign ups from 21 February 2025 and moving to turn it off for existing UK users. Apple said it was gravely disappointed, that it hopes to offer the feature in the UK again, and that it has never created a backdoor or master key for any of its products.&lt;/p&gt;
&lt;p&gt;That story cuts 2 ways. The strong protection is real, when Advanced Data Protection is on, the data is out of Apple’s reach and therefore out of a government’s reach through Apple. A government with enough power can still force the protection to be withdrawn in its territory.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-04.50y8bkO-_klTwc.webp&quot; alt=&quot;Figure 4. The iCloud encryption fight, start to finish. Apple added end to end encrypted Advanced Data Protection in December 2022, the UK secretly ordered a worldwide backdoor in early 2025, and Apple pulled the feature for UK users from 21 February 2025 rather than build one. It shows both that the strong protection is real and that a government can force it offline locally.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1088&quot; height=&quot;544&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Figure 4. The iCloud encryption fight, start to finish. Apple added end to end encrypted Advanced Data Protection in December 2022, the UK secretly ordered a worldwide backdoor in early 2025, and Apple pulled the feature for UK users from 21 February 2025 rather than build one. It shows both that the strong protection is real and that a government can force it offline locally. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.bleepingcomputer.com/news/security/apple-pulls-icloud-end-to-end-encryption-feature-in-the-uk%20It%20is%20the%20clearest%20illustration%20that%20the%20hardest%20privacy%20guarantee%20is%20encryption%20only%20you%20hold%20the%20key%20to,%20and%20that%20even%20that%20runs%20into%20the%20limits%20of%20where%20you%20live%20and%20what%20laws%20apply.%20Windows%20and%20OneDrive%20have%20no%20exact%20equivalent%20to%20end%20to%20end%20encrypted%20cloud%20backup%20that%20the%20vendor%20cannot%20read,%20which%20is%20a%20real%20gap%20on%20the%20Microsoft%20side%20for%20anyone%20who%20backs%20up%20to%20the%20cloud.&quot;&gt;Author timeline based on reporting and Apple statements&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;the-small-signals-that-tell-you-when-you-are-being-watched&quot;&gt;The small signals that tell you when you are being watched&lt;/h2&gt;
&lt;p&gt;Some of the most useful privacy features are the ones that just tell you what is happening. On a Mac the camera has a green light wired into the same circuit that powers it, so the light physically comes on whenever the camera is recording, and software cannot switch the camera on while leaving the light off. macOS also shows a dot in the menu bar, orange when the microphone is active and green for the camera, so an app cannot quietly listen or watch without a visible tell. Windows shows a camera indicator too, but the Mac camera light being wired to the hardware rather than drawn by software is a stronger guarantee.&lt;/p&gt;
&lt;p&gt;Apple also gives you 2 network level tools worth turning on. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://support.apple.com/guide/icloud/mm8010d8daf3/icloud&quot;&gt;iCloud Private Relay&lt;/a&gt;, included with a paid iCloud+ plan, routes your Safari traffic through 2 separate relays so that no single party sees both who you are and which site you are visiting, Apple sees your IP but not the destination, a partner sees the destination but not your IP. It is narrower than a full VPN, it only covers Safari and does not let you pick a country, but it hides your address from the sites and networks you touch. The second is private Wi-Fi addressing, which rotates the hardware identifier your device broadcasts on each network so you are harder to track as you move between them. Windows has no built in equivalent to either, you would reach for a third party VPN and manual settings instead.&lt;/p&gt;
&lt;h2 id=&quot;how-much-work-it-takes-to-lock-each-one-down&quot;&gt;How much work it takes to lock each one down&lt;/h2&gt;
&lt;p&gt;Settings only help if you can find them. I went through the realistic checklist for each system, the things a privacy minded person would actually turn off.&lt;/p&gt;
&lt;p&gt;On Windows 11 the list is long and spread out. Turn off the advertising ID, cut optional diagnostic data, check location permissions, review camera and microphone access, look at inking and typing personalization, audit background apps for anything acting like a keylogger, check OneDrive syncing, review Edge’s privacy and sync settings, see which apps can read your account info, open the privacy dashboard, open the diagnostic data viewer, confirm encryption is on, and if you have Recall, turn it off and delete any existing snapshots. It is thorough because it has to be, and it is scattered across many screens.&lt;/p&gt;
&lt;p&gt;On macOS the equivalent pass is shorter and more centralized. Turn off analytics sharing with Apple and with developers, disable Siri and dictation if you want, turn off ad personalization, check location services, camera, microphone, and screen recording permissions, review full disk access and accessibility, check startup and background apps, review iCloud syncing, check Safari’s privacy settings, enable FileVault, and review the devices attached to your Apple account. Fewer stops, and they live closer together.&lt;/p&gt;
&lt;p&gt;There is a catch on the Windows side that users report often, that a major update can quietly flip some privacy settings back toward the defaults, so the pass you did last year may need doing again. It is common enough that a small industry of tools like O&amp;amp;O ShutUp10 exists mainly to reapply the same settings after each update. Treat the Windows lockdown as something to re check rather than set once.&lt;/p&gt;
&lt;p&gt;Neither system is simply secure or insecure. What separates them is that macOS makes the private configuration easier to reach, and easier settings get used. A privacy control buried 5 menus deep is one most people never touch.&lt;/p&gt;
&lt;h2 id=&quot;the-honest-verdict-and-the-asterisk-on-all-of-it&quot;&gt;The honest verdict, and the asterisk on all of it&lt;/h2&gt;
&lt;p&gt;On the comparison most people actually face, a Mac out of the box leans more private than a Windows 11 PC out of the box, mostly because of defaults and discoverability. Apple’s business does not depend on harvesting your behavior the way an ad and AI data strategy does, the privacy settings are easier to find and use, FileVault and Lockdown Mode are strong, and features like Siri and Apple Intelligence are genuinely optional. Windows can be locked down to a similar place, but it takes more work and more knowledge, and some defaults have started in the wrong spot.&lt;/p&gt;
&lt;p&gt;None of that makes either one the most private option available. Both send telemetry home, both comply with lawful government requests, and both are run by companies with commercial reasons to know things about you. If privacy is your absolute priority, this road ends at Linux, a free and open operating system built from many distributions like Debian, Ubuntu, and Fedora. Linux is not bulletproof either, it is assembled from many independent pieces, any of which can carry a vulnerability for days or weeks, so it demands that you keep it updated and pay attention to what is running. The payoff is a system nobody is monetizing and that you fully control. For most people that is more maintenance than they want, which is exactly why the real choice stays Windows versus Mac, and why knowing what each actually does with your data is worth the half hour it takes to set up.&lt;/p&gt;</content:encoded><dc:creator>Shivashish Yadav</dc:creator><category>Security</category><category>mac</category><category>macos</category><category>microsoft</category><category>ai</category><category>data</category><author>hello@soursen.org (Shivashish Yadav)</author></item><item><title>The Arch AUR Keeps Getting Hit With Malware, and the Fixes All Cost Something</title><link>https://soursen.org/articles/the-arch-aur-keeps-getting-hit-with-malware-and-the-fixes-all-cost/</link><guid isPermaLink="true">https://soursen.org/articles/the-arch-aur-keeps-getting-hit-with-malware-and-the-fixes-all-cost/</guid><description>On 13 October 2026 an account mass uploaded a pile of packages to the Arch User Repository that looked like the popular helper yay and ran code the attacker controlled as root the moment you installed them. It was the fourth malware wave on the AUR in about 15 months. I went through what actually happened, how the trick worked at the file level, and why the obvious fix is the one thing that would stop the AUR from being the AUR.</description><pubDate>Mon, 26 Oct 2026 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;If you run Arch Linux, you already know the &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://aur.archlinux.org/&quot;&gt;AUR&lt;/a&gt;, the Arch User Repository. It is the reason people put up with Arch, a giant community library where anyone can publish a build recipe for software that is not in the official repos. That openness is the whole appeal, and it is also the problem, because anyone includes people uploading malware. Over roughly the last 15 months the AUR has been hit four separate times, and the most recent one, in October 2026, was the crudest and in some ways the most telling.&lt;/p&gt;
&lt;p&gt;I want to walk through what happened, because the mechanism is simple enough to understand in a few minutes and it explains why this keeps happening. None of the fixes are free, and that is the uncomfortable part.&lt;/p&gt;
&lt;h2 id=&quot;what-an-aur-package-actually-is-and-why-that-matters&quot;&gt;What an AUR package actually is, and why that matters&lt;/h2&gt;
&lt;p&gt;A package in the official Arch repos is built and signed by a trusted Arch maintainer. A package in the AUR is a text file called a PKGBUILD that some random person uploaded. When you install an AUR package with a helper like yay or paru, that helper downloads the PKGBUILD, runs the build steps on your machine, and installs the result. PKGBUILDs can also carry install scripts that run at install or upgrade time, and those scripts run as root, because installing software needs root.&lt;/p&gt;
&lt;p&gt;Read that again, because it is the entire attack surface. You are downloading a script written by a stranger and running part of it as root on your own machine. The AUR has always told you to read the PKGBUILD before installing. Almost nobody does. That gap between what you are supposed to do and what you actually do is what every one of these attacks lives in.&lt;/p&gt;
&lt;h2 id=&quot;the-october-2026-wave-a-plain-old-typosquat&quot;&gt;The October 2026 wave, a plain old typosquat&lt;/h2&gt;
&lt;p&gt;On 13 October 2026, an account called &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.bleepingcomputer.com/news/security/fake-arch-linux-packages-push-malware-with-root-privileges/&quot;&gt;marchos1as mass uploaded packages&lt;/a&gt; that impersonated popular AUR helpers. The mailing list thread that raised the alarm was titled, bluntly, hundreds of packages mass uploaded with malware. The names were typosquats of real tools, the clearest example being yyay, an extra y bolted onto yay, the single most used AUR helper there is.&lt;/p&gt;
&lt;p&gt;Typosquatting is the oldest trick in the book. You register a name that is a near miss for a popular one and wait for people to fat finger it. PayPal has spent years buying up domains like paypa1 and paypai so that a mistyped address lands back on the real site. On the AUR this has not historically been a thing people worried about, which is exactly why it worked. Most of the fake names were nonsense like Hay, Yeah, and Yizzy, the kind of thing nobody would ever type by accident. The point was volume. Upload enough misspellings and a few people will eventually install one.&lt;/p&gt;
&lt;p&gt;Here is what made it dangerous rather than just annoying. The fake yyay package &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.helpnetsecurity.com/2026/10/16/arch-linux-aur-malware-yay-paru/&quot;&gt;presented itself as a straightforward repackaging of yay&lt;/a&gt;, same upstream URL, same source table, so a glance at it looked fine. The install script, yyay.install, is where the payload sat. The post install and post upgrade functions were deliberately left empty, filled with a harmless true, so an automated scanner or a casual reviewer skimming for the obvious hook spots would see nothing. The actual malicious line sat at the top level of the script file, outside any function, which means it runs unconditionally as root on every install and upgrade.&lt;/p&gt;
&lt;p&gt;That line piped a remote shell script straight into a shell. It pulled a file called linux.sh from a personal project on gitflic.ru, a Russian code host with no connection to the real yay project, and it wrapped the URL in a web.archive.org link that did not pin a snapshot timestamp. No pinned timestamp means the content served could change at any time, and there was no checksum on the fetch, so whoever controlled that script could swap the payload for credential theft one day and a rootkit the next. You would never know which one you got.&lt;/p&gt;
&lt;p&gt;There is a thread worth pulling here. The repository the script pointed at, named something like hellscripter and install-scripts, matches the setup used in a separate npm campaign that researchers tracked under the name &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.scworld.com/brief/malicious-packages-impersonating-arch-linux-tools-found-in-aur&quot;&gt;DirtyBlanket&lt;/a&gt;, which pulled code from a similar hellscripter repository and deployed the CHAOS remote access trojan. Nobody has confirmed the two operations are the same crew, and I am not going to claim they are. What the overlap tells you is that these AUR attacks are not isolated pranks, they sit inside a wider pattern of people planting the same kind of install time payloads across multiple package ecosystems.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-01.DyHm5qVK_Z2cDoC.webp&quot; alt=&quot;Figure 1. How the fake yyay package hid its payload. The install hooks were emptied with true so scanners saw nothing, while the real command sat at the top level of the script and ran as root, piping a remote linux.sh from a Russian host straight into a shell with no checksum and no pinned version.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1049&quot; height=&quot;627&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Figure 1. How the fake yyay package hid its payload. The install hooks were emptied with true so scanners saw nothing, while the real command sat at the top level of the script and ran as root, piping a remote linux.sh from a Russian host straight into a shell with no checksum and no pinned version. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.helpnetsecurity.com/2026/10/16/arch-linux-aur-malware-yay-paru/&quot;&gt;Author diagram based on the published malware scan&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;the-one-genuinely-good-part-it-got-caught-fast&quot;&gt;The one genuinely good part, it got caught fast&lt;/h2&gt;
&lt;p&gt;Credit where it is due. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.helpnetsecurity.com/2026/10/16/arch-linux-aur-malware-yay-paru/&quot;&gt;Less than an hour after the report&lt;/a&gt; hit the mailing list, an Arch maintainer confirmed the malicious payloads, banned the marchos1as account, and deleted every package it had uploaded. A maintainer running a detection tool called TrustSight helped flag the bad packages. So the live window was short, and most of the nonsense names almost certainly never got installed by anyone, since you need to get at least the first few characters of a command right before a typo even becomes plausible.&lt;/p&gt;
&lt;p&gt;That is the good news and it is real. The bad news is that fast cleanup is a treadmill, not a fix. The account got malware onto the AUR during a period when the project had already locked things down because of earlier attacks, which tells you the lockdown is porous.&lt;/p&gt;
&lt;h2 id=&quot;this-was-the-fourth-wave-not-the-first&quot;&gt;This was the fourth wave, not the first&lt;/h2&gt;
&lt;p&gt;The October typosquat did not come out of nowhere. The AUR has been a repeat target.&lt;/p&gt;
&lt;p&gt;In July 2025, a user called danikpapas uploaded three packages named &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://itsfoss.com/news/arch-linux-chaos-rat/&quot;&gt;firefox-patch-bin, librewolf-fix-bin, and zen-browser-patched-bin&lt;/a&gt;. They looked like browser fixes. They installed CHAOS RAT, a remote access trojan that hands an attacker control of the machine. They went up on 16 July and were removed on 18 July, so they were live for about 2 days.&lt;/p&gt;
&lt;p&gt;In the week of 12 June 2026, the attack was far more sophisticated, and it is worth sitting with because it shows how bad this can get. Attackers adopted orphaned AUR packages, ones whose maintainers had walked away, then edited the PKGBUILD or .install scripts so that building the package quietly ran npm install atomic-lockfile, with a couple of real npm packages alongside it for cover. They even &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://thehackernews.com/2026/06/over-400-arch-linux-aur-packages.html&quot;&gt;forged the git commit metadata&lt;/a&gt; so the changes looked like they came from the long standing maintainers, though an Arch Trusted User confirmed those accounts were not actually compromised. A community count using grep reached about 408 affected packages, and the list was described as still growing and incomplete.&lt;/p&gt;
&lt;p&gt;The payload was nasty. The atomic-lockfile npm package carried a preinstall hook that ran a bundled Linux binary, a Rust credential stealer aimed squarely at developer machines. It went after cookies, tokens, and local storage from Chromium browsers, logged in sessions from Electron apps like Slack, Discord, and Teams, GitHub and npm and HashiCorp Vault tokens, OpenAI bearer material, SSH keys and known_hosts, shell history, Docker and Podman credentials, and VPN profiles. It shipped the loot to temp.sh over plain HTTP and ran its command and control over Tor. If the binary already had root, it could load an optional eBPF rootkit that hid its own processes and sockets and blocked debuggers, and it even staged a file tied to monero-wallet-gui that looked like an unanalyzed cryptominer. Anyone who installed or updated an affected package on or after 11 June should assume those secrets are gone.&lt;/p&gt;
&lt;p&gt;In late July 2026, another wave pushed Arch to suspend package adoption, the exact mechanism that let the June attackers grab orphaned packages, and within days the project halted new AUR uploads and account registrations entirely. If you go to the AUR register page today, you still hit a notice that new account registration is temporarily closed. That lockdown has held for more than a month, and it is the backdrop to the October incident, because marchos1as managed to upload malware anyway, which means the account already existed before the freeze.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-02.DUtHJLdg_Z2etAKG.webp&quot; alt=&quot;Figure 2. Four AUR malware waves in about 15 months, from the July 2025 CHAOS RAT browser packages to the June 2026 takeover of more than 400 orphaned packages with a Rust infostealer, the July 2026 upload freeze, and the October 2026 typosquatting run.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1029&quot; height=&quot;569&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Figure 2. Four AUR malware waves in about 15 months, from the July 2025 CHAOS RAT browser packages to the June 2026 takeover of more than 400 orphaned packages with a Rust infostealer, the July 2026 upload freeze, and the October 2026 typosquatting run. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://thehackernews.com/2026/06/over-400-arch-linux-aur-packages.html&quot;&gt;Author timeline based on Arch mailing list reports and security coverage&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;the-sleeper-account-problem-nobody-wants-to-talk-about&quot;&gt;The sleeper account problem nobody wants to talk about&lt;/h2&gt;
&lt;p&gt;The October wave points at something uncomfortable. Registrations were closed. Uploads were supposedly restricted. The account still pushed malware onto the AUR. That lines up with a worry that was raised during the earlier waves, that there are dormant accounts sitting on the AUR, created before the freeze, that have done nothing yet and are waiting. Closing the front door does nothing about the people already inside the house.&lt;/p&gt;
&lt;p&gt;I have no way to confirm how many such accounts exist, and I would not trust anyone who claims a precise number. What I can say is that the pattern fits. An account that predates a registration freeze, surfacing mid lockdown to dump hundreds of typosquats, is exactly what a sleeper account would look like. The motive, based on what has been reported from people claiming involvement in these campaigns, is not some grand cause. It is that they can. When the only thing standing between an attacker and a root shell on your machine is whether you bothered to read a text file, and the payoff is real, someone will keep trying.&lt;/p&gt;
&lt;h2 id=&quot;the-other-weird-thing-happening-on-the-mailing-list&quot;&gt;The other weird thing happening on the mailing list&lt;/h2&gt;
&lt;p&gt;Running alongside the malware is a smaller, stranger problem that points in the same direction. The AUR general mailing list has been getting waves of near identical emails that just say leave, or I’d like to unsubscribe from this mailing list, often several on the same day and then nothing. The sender posts once and vanishes, never replies, never explains. Leaving the list is trivial, you email a dedicated address, aur-general-leave at the list domain, so a human who wanted out would manage it. Dozens of people suddenly unable to figure that out, in a month where it almost never happened before, does not read like dozens of confused humans.&lt;/p&gt;
&lt;p&gt;The likely explanation is AI agents. Automated agents subscribed to the list and now cannot work out how to unsubscribe, so they keep firing the same request into the void. One of them even threatened to report Arch Linux to Hetzner, the hosting provider, for sending unsolicited spam, which, if Hetzner took it seriously, could get Arch’s mail relay port throttled while they investigated, a real headache from a complaint that makes no sense. It is low stakes on its own, maybe 20 emails over a month, annoying rather than damaging. It matters because it is the same root cause as the malware, automated accounts interacting with the AUR’s open infrastructure faster than humans can reason about them, and nobody being quite sure which messages are bots and which are people.&lt;/p&gt;
&lt;h2 id=&quot;why-the-obvious-fix-breaks-the-thing-you-are-trying-to-fix&quot;&gt;Why the obvious fix breaks the thing you are trying to fix&lt;/h2&gt;
&lt;p&gt;So how do you stop this? The clean answer is to treat the AUR like the official repos. The official Arch repos almost never ship malware, because only a small set of trusted maintainers can put packages in them, and everything is reviewed and signed. Debian works the same way. Lock the AUR down to verified contributors, review submissions, sign the results, and the typosquatting problem mostly evaporates.&lt;/p&gt;
&lt;p&gt;Except that destroys the AUR. The entire reason the AUR exists is that anyone can publish anything without waiting for a gatekeeper. The moment you add trusted maintainers and a review queue, you have built a second set of official repos, slower and smaller, and the open community library that made Arch worth using is gone. You would have solved the malware problem by deleting the feature.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-03.DphB5sQn_Z1fYas9.webp&quot; alt=&quot;Figure 3. The tradeoff behind every proposed fix. Official repos are safe because they are closed and reviewed. The AUR is useful because it is open to anyone, which is also why it keeps getting hit. Locking it down would stop the malware and stop it being the AUR.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1040&quot; height=&quot;601&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Figure 3. The tradeoff behind every proposed fix. Official repos are safe because they are closed and reviewed. The AUR is useful because it is open to anyone, which is also why it keeps getting hit. Locking it down would stop the malware and stop it being the AUR. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://aur.archlinux.org/&quot;&gt;Author diagram&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;p&gt;The realistic options are all compromises. Keep registrations locked and reopen them one verified request at a time, which is slow and manual and does not touch the sleeper accounts. Add automated scanning of PKGBUILDs, which the October attack was specifically built to evade by hiding the payload outside the functions scanners check. Lean on third party detection tools like the TrustSight scanner maintainers have started running, which help but arrive after upload. Or shift the burden back onto users and actually get people to read PKGBUILDs before installing, which is the official advice and has never worked at scale because convenience always wins.&lt;/p&gt;
&lt;p&gt;None of these is a solution. They are all ways of making the attack a little harder or a little slower while keeping the AUR open enough to still be the AUR. That tension does not resolve. As long as the AUR lets strangers publish code that runs as root on your machine, and as long as most people install without reading, the AUR will keep getting hit, and the response will keep being fast cleanup rather than prevention.&lt;/p&gt;
&lt;p&gt;If you use the AUR, the practical takeaway is small and boring and worth doing anyway. Read the PKGBUILD and the .install file before you install, every time, especially for anything you typed quickly. Check that the package name is spelled exactly right. Be suspicious of any package that fetches a script from a host that has nothing to do with the real project. You can check what is installed with pacman -Qs and remove something with sudo pacman -Rns, though if you ran a root payload, removing the package does not undo what it already did, and a clean reinstall from backup is the safer call.&lt;/p&gt;</content:encoded><dc:creator>Shivashish Yadav</dc:creator><category>Security</category><category>linux</category><category>arch</category><category>aur</category><category>malware</category><category>security</category><author>hello@soursen.org (Shivashish Yadav)</author></item><item><title>eMudhra’s CA Keys Were Compromised Because a Private Key Was the Number 4.</title><link>https://soursen.org/articles/emudhras-ca-keys-were-compromised-because-a-private-key-was-the-number-4/</link><guid isPermaLink="true">https://soursen.org/articles/emudhras-ca-keys-were-compromised-because-a-private-key-was-the-number-4/</guid><description>In October 2026, Mozilla and Google confirmed that 8 intermediate signing keys from eMudhra, 1 of the few Indian certificate authorities trusted globally, had private keys a laptop can guess in under a second, including 1 that was literally the integer 4. I went through how a secret that should be 116 digits long collapsed to single digits, how browsers shut it down in under 3 hours, why your backend servers may still be exposed, and why, for a certificate authority, the incident is survivable but the response decides everything.</description><pubDate>Thu, 22 Oct 2026 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;The padlock in your browser rests on a single assumption, that a certain secret number is so enormous no one could ever guess it. In October 2026 that assumption broke in the most embarrassing way possible. A security researcher found that 8 intermediate signing keys belonging to eMudhra, 1 of India’s largest certificate authorities, had private keys small enough to recover on a laptop in under a second. One of them was the number 4. Not a 116 digit number starting with 4. The number 4.&lt;/p&gt;
&lt;p&gt;This is the kind of failure that sounds made up, and it is fully documented in the public record. I want to walk through exactly what happened, because the mechanism is a clean lesson in how internet trust works, how it fails, and why the fix reached your browser almost instantly while leaving a large part of the internet’s plumbing exposed.&lt;/p&gt;
&lt;h2 id=&quot;what-a-certificate-authority-actually-does&quot;&gt;What a certificate authority actually does&lt;/h2&gt;
&lt;p&gt;Every time you load a site over HTTPS, your browser has to answer 1 question, is this really the server it claims to be. The answer comes from public key cryptography. The server holds a private key it never shares, and publishes a matching public key inside a certificate. The server proves its identity by signing data with the private key, and your browser checks that signature with the public key. Anyone can see the public key. Only the holder of the private key can produce a valid signature, so a valid signature is proof.&lt;/p&gt;
&lt;p&gt;The missing piece is, why should your browser trust that a given public key really belongs to your bank. That is the job of a certificate authority, a trusted notary that signs the server’s certificate with its own private key, binding the public key to a verified identity. Your devices ship with a built in list of these authorities, called a trust store, maintained by the big root programs at &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://blog.mozilla.org/security/2026/10/20/protecting-firefox-users-following-the-compromise-of-emudhra-intermediate-certificates/&quot;&gt;Mozilla&lt;/a&gt;, Apple, Microsoft, and Google. Before you visit a single website, your device already trusts every authority on that list.&lt;/p&gt;
&lt;p&gt;The authorities do not sign websites directly with their crown jewels. At the top sits a root key, kept offline in a hardware vault and switched on only for rare, formal ceremonies, because if a root key ever leaked the only fix is to remove it from every device on Earth. Below the root sit intermediate keys, which do the daily work of signing website certificates while staying online. If an intermediate goes bad, the root can revoke it and sign a new 1 without anyone’s device changing a thing. That design is the only reason this story did not end in total disaster.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-01.DCoNhTN2_1wMlIO.webp&quot; alt=&quot;Figure 1. The chain of trust. A root certificate anchored in your device signs an intermediate, the intermediate signs individual website certificates, and your browser checks each signature up the chain. eMudhra’s compromised keys were intermediates, which can issue certificates for any domain.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1132&quot; height=&quot;709&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Figure 1. The chain of trust. A root certificate anchored in your device signs an intermediate, the intermediate signs individual website certificates, and your browser checks each signature up the chain. eMudhra’s compromised keys were intermediates, which can issue certificates for any domain. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/Public_key_infrastructure&quot;&gt;Author diagram of public key infrastructure&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;the-number-that-should-have-been-116-digits-long&quot;&gt;The number that should have been 116 digits long&lt;/h2&gt;
&lt;p&gt;The 8 compromised keys used an elliptic curve called P-384. On that curve a private key is supposed to be a random integer up to roughly 10 to the power 115, a number about 116 digits long. The security of the whole system depends on that number being genuinely random and genuinely enormous, so that guessing it would take longer than the lifetime of the universe.&lt;/p&gt;
&lt;p&gt;eMudhra’s keys were not enormous. Per Mozilla’s incident report, all 8 had values below 256, meaning the entire secret fit in a single byte. A researcher named &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.helpnetsecurity.com/2026/10/21/emudhra-certificates-private-keys-compromised/&quot;&gt;Scott Helme&lt;/a&gt; recovered them from the public certificates on a laptop, and said it took less than a second, because there are only 256 possibilities to try. The actual values were tiny integers like 4, 56, and 115. The key for the intermediate labeled EM OV TLS CA — G3C had a value of 4, confirmed by BleepingComputer against the revoked certificate. A 4 digit bike lock has more combinations than that key did.&lt;/p&gt;
&lt;p&gt;The funny part has a catastrophic edge. These were intermediate keys, which means whoever holds them can mint a valid certificate for any domain they like, your bank, your email, a government portal, and every system checking the chain would report it as completely legitimate. A stolen intermediate key is a passport printing machine for the entire internet. The only saving grace is that using it to steal your login still requires an attacker to sit on your network and intercept your traffic, so the keys being public did not automatically drain anyone’s bank account. It did mean the single assumption the whole padlock rests on was false for 17 months.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-02.U-rNg-9s_1mQwaA.webp&quot; alt=&quot;Figure 2. The gap between the key eMudhra should have generated and the key it used. A P-384 private key should be about 116 digits long. The 8 compromised keys were integers below 256, 1 of them the number 4, recoverable in under a second.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1182&quot; height=&quot;603&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Figure 2. The gap between the key eMudhra should have generated and the key it used. A P-384 private key should be about 116 digits long. The 8 compromised keys were integers below 256, 1 of them the number 4, recoverable in under a second. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.helpnetsecurity.com/2026/10/21/emudhra-certificates-private-keys-compromised/&quot;&gt;Author chart based on Mozilla and researcher reports&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;how-a-key-ceremony-produced-a-4&quot;&gt;How a key ceremony produced a 4&lt;/h2&gt;
&lt;p&gt;Certificate authorities generate keys in what is called a key ceremony, one of the most tightly controlled procedures a security company runs. It happens in a locked room, phones left outside, an offline computer wired to a tamper resistant hardware security module, several people each holding a smart card so nobody can act alone, a written script read step by step, and an auditor recording the whole thing on camera. The entire ritual exists to guarantee 1 outcome, that the right keys are generated and the right keys get signed.&lt;/p&gt;
&lt;p&gt;eMudhra held such a ceremony on 30 April 2025 and generated 24 intermediates in a single sitting. Per &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://emudhra.com/en/post/certificate-security-incident&quot;&gt;the company’s own account&lt;/a&gt;, a flawed test script created a second, unintended set of keys alongside the proper ones, and those second keys were the weak ones. Because the weak keys carried labels nearly identical to the strong keys, staff selected the wrong ones and built certificate requests from them, and those got signed and published into the global browser trust stores. The hardware vault worked perfectly. It generated proper random keys exactly as designed. A procedural script running next to it quietly produced garbage, and a human picked the garbage by mistake.&lt;/p&gt;
&lt;p&gt;There was a second chance to catch it, and it got missed in a way that is hard to read without wincing. When a certificate authority wants to revoke a certificate, it publishes a signed revocation list. At one point eMudhra signed such a list with the real, strong key, and it failed to verify against the certificate, because the certificate in use was actually built on the weak key. That mismatch was the system screaming that something was wrong. The reported response was to re sign the list with the weak key so it would match, when the mismatch should have stopped everything while someone asked why the keys did not line up. The weak key stayed in use.&lt;/p&gt;
&lt;h2 id=&quot;the-internet-shut-it-down-in-under-3-hours&quot;&gt;The internet shut it down in under 3 hours&lt;/h2&gt;
&lt;p&gt;The discovery itself reads like a thriller. On Saturday 17 October 2026, a researcher credited by Mozilla only as Wayne was scanning public certificate data, the append only logs where authorities must publish every certificate they issue. Finding the weak keys, Wayne emailed the root programs and signed the message, we have a problem here, with each of the 8 compromised keys. Signing a message with a certificate authority’s own keys is the most polite imaginable way of saying, I have your private keys, and here is the proof.&lt;/p&gt;
&lt;p&gt;What happened next is the part that should give you some confidence in the modern web. The timeline, in UTC, is tight. At 18:05 the Chrome Root Program received the report. By 19:45, about 1 hour and 40 minutes later, Chrome’s block was live, pushed to browsers through a mechanism called a CRLSet that needs no update or restart from the user. Mozilla received the report at 20:06 and finished pushing its own block through OneCRL by 22:59, under 3 hours. Apple added protections on 19 October, and Microsoft on 20 October. The browser makers blocked the keys themselves at the key level, which reaches further than blocking individual certificates, and that matters because an older authority called Sectigo had cross signed the same keys under a different root. Blocking the key catches every copy at once.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-03.B0v_oPTO_Z1Yj098.webp&quot; alt=&quot;Figure 3. From report to global block in under 3 hours. The response timeline on 17 October 2026, in UTC, from the researcher’s email to Chrome and Firefox pushing blocks to billions of devices without a user update.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1165&quot; height=&quot;480&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Figure 3. From report to global block in under 3 hours. The response timeline on 17 October 2026, in UTC, from the researcher’s email to Chrome and Firefox pushing blocks to billions of devices without a user update. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://thehackernews.com/2026/10/eight-emudhra-ca-keys-compromised.html&quot;&gt;Author timeline based on Chrome and Mozilla incident reports&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;your-browser-is-safe-your-servers-might-not-be&quot;&gt;Your browser is safe, your servers might not be&lt;/h2&gt;
&lt;p&gt;Browsers solved this fast because they have a central kill switch, a block list pushed silently to billions of devices in the background. The trouble is that a huge amount of the internet does not run in a browser. Headless Linux servers, microservices, containers, and automated pipelines validate certificates against a static trust store sitting on disk, a file like the ca-certificates package on Linux or a bundle copied into a Docker image years ago. These systems never fetch the browser block lists. They trust whatever file is on disk until a human updates it.&lt;/p&gt;
&lt;p&gt;That is the real exposure. If a backend service makes an outbound connection and its local trust store still lists the eMudhra root, standard tools like curl on Linux or macOS will happily verify a forged certificate as perfectly valid. Mozilla’s Dennis Jackson specifically warned that the built in curl on macOS may keep trusting the certificates, because its bundle does not check revocation. Pipelines using mutual TLS to authenticate server to server traffic, with no human in the loop, are exactly the places this lingers. Browser updates protect people quickly and offer zero coverage for this machinery until an administrator intervenes.&lt;/p&gt;
&lt;p&gt;The cleanup reflects that split. Mozilla removed the root entirely from its NSS trust store in version 3.131.1, Go shipped versions 1.27.3 and 1.26.10 that reject any chain containing 1 of the 8 keys under CVE-2026–111943, and bundles like certifi, Rust’s webpki-roots, and Homebrew’s ca-certificates pulled the root as well. Each of those is a separate update on a separate schedule, and anything that does not pull them stays exposed. eMudhra estimates about 30,000 customer certificates need to be replaced, and has said it will recheck every key ceremony it has ever carried out.&lt;/p&gt;
&lt;h2 id=&quot;trust-is-the-whole-product-and-it-is-the-hard-part-to-get-back&quot;&gt;Trust is the whole product, and it is the hard part to get back&lt;/h2&gt;
&lt;p&gt;A certificate authority sells exactly 1 thing, trust. Not software, not storage, the assurance that when its signature says this is your bank, it really is. When the signing key behind that assurance turns out to be the number 4, the product itself is what failed, and that is why this incident lands harder than a normal bug. The reaction across the Indian security community has carried a note of personal disappointment, because eMudhra is 1 of the few Indian certificate authorities trusted in the global browser root stores, a genuine point of pride, and that standing is now under a spotlight it did not want.&lt;/p&gt;
&lt;p&gt;The uncomfortable history here is that in the certificate authority world, a technical mistake is usually survivable while a mishandled response usually is fatal. DigiNotar was breached in 2011, handled it poorly, got pulled from the trust stores, and was bankrupt within weeks. Symantec, once a giant of the industry, lost the trust of Google and Mozilla over a pattern of mis-issuance and had its certificate business wound down. Entrust was distrusted by Chrome and Firefox in 2024 after a run of incidents and what the browsers judged to be inadequate responses. The browsers do not remove a CA for a single flaw. They remove it when they lose confidence that the CA will catch the next one, which is a judgment about transparency and competence that reaches well beyond this specific key.&lt;/p&gt;
&lt;p&gt;That is why the next few weeks matter more than the keys themselves. What the industry is watching for is a full, public root cause analysis that answers the questions that actually determine trust. How did a script generate a single digit key in the first place. Why did the near-identical labels let a human pick the weak one. Why did the certificate revocation list mismatch, the moment the system flagged that something was wrong, get papered over when it should have been escalated. How did these certificates sit in public logs for 17 months with nobody, at eMudhra or at the cross-signing partner Sectigo, running the 1 second check that would have caught them. What changes now so it cannot recur. eMudhra has said it will recheck every key ceremony it has ever performed, which is the right instinct. The detail and candor of what it publishes next is what the root programs will weigh.&lt;/p&gt;
&lt;p&gt;For everyone else in the ecosystem, there is a quieter lesson. A single CA’s mistake lands on people who have never heard of that CA, because trust in this system is inherited. Your browser, your language runtime, and your Docker image all pull their trusted roots from lists maintained by a handful of organizations, and a weak key signed by 1 authority becomes everyone’s problem on whatever day their particular list gets updated. The resilience of the web came not from eMudhra being perfect, but from the layers above it, certificate transparency logs that made the keys findable, and browser kill switches that could act in hours.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-04.B7z8ZWSk_Z2o3lT8.webp&quot; alt=&quot;Figure 4. For a certificate authority, surviving an incident is about the response. Past cases show browsers remove a CA not for a single flaw, but when they lose confidence in how it responds. DigiNotar, Symantec, and Entrust all lost trust over how they handled their incidents, beyond the original mistakes.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1296&quot; height=&quot;544&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Figure 4. For a certificate authority, surviving an incident is about the response. Past cases show browsers remove a CA not for a single flaw, but when they lose confidence in how it responds. DigiNotar, Symantec, and Entrust all lost trust over how they handled their incidents, beyond the original mistakes. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.encryptionconsulting.com/firefoxs-mozilla-follows-google-in-distrusting-entrusts-tls-certificates/&quot;&gt;Author summary of certificate authority trust history&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;what-to-actually-do-if-you-run-infrastructure&quot;&gt;What to actually do if you run infrastructure&lt;/h2&gt;
&lt;p&gt;If you only use a browser, update it and move on, the block is already there. If you run servers, the work is manual and worth doing now. Audit your container images for outdated certificate bundles, since a base image built a couple of years ago may still carry the eMudhra root. Run your operating system’s package manager to force the local trust store to reflect current revocation data, and purge any certificate tied to the eMudhra G3 chain from your directories. If your own site’s certificate chains to the affected root, get a replacement from a different authority now.&lt;/p&gt;
&lt;p&gt;The 1 thing not to do is reach for the quick fix when connections start failing mid migration. Disabling certificate validation with a flag like verify equals false makes the errors go away and leaves a permanent hole that trusts everything, forged certificates included. Configure your continuous integration to actively reject any chain built on a revoked root instead.&lt;/p&gt;
&lt;p&gt;The eMudhra failure is a reminder that internet security is not an automatic guarantee handed down by browser vendors. The math behind it never broke. A certified hardware vault did its job. The trust still collapsed, because a script produced a single digit number, a human picked the wrong label, a mismatch got papered over, and for 17 months a perfectly valid looking certificate hid a private key of 4 that nobody checked.&lt;/p&gt;</content:encoded><dc:creator>Shivashish Yadav</dc:creator><category>Security</category><category>security</category><category>cybersecurity</category><category>emudhra</category><category>mozilla</category><category>firefox</category><author>hello@soursen.org (Shivashish Yadav)</author></item><item><title>Anthropic’s New AI Threat Report Is One of the Strangest Documents in Tech</title><link>https://soursen.org/articles/anthropics-new-ai-threat-report-is-one-of-the-strangest-documents-in/</link><guid isPermaLink="true">https://soursen.org/articles/anthropics-new-ai-threat-report-is-one-of-the-strangest-documents-in/</guid><description>Chinese labs ran nearly 200 million exchanges through Claude. A weapons cell in Yemen used it as an engineering team. Here is what holds up, and where the skepticism is fair.</description><pubDate>Mon, 21 Sep 2026 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;2 headlines landed from the biggest US AI labs inside 48 hours, and they could not have been more different in mood. OpenAI said a swarm of roughly 10,000 agents produced a machine verified result on the Navier Stokes problem, one of the 7 Millennium Prize problems, in 88 hours. Anthropic, meanwhile, published a 154 page document cataloguing the people who tried to use its product to hurt other people.&lt;/p&gt;
&lt;p&gt;I read the whole thing. It is one of the strangest and most specific documents a major AI company has ever put out, and I want to walk you through what is actually in it, what holds up, and where I think the skepticism is fair.&lt;/p&gt;
&lt;p&gt;One caution before any of the details, and it applies to everything below. Anthropic published this report on September 10, 2026. Every claim in it is Anthropic’s own account, investigated by Anthropic, described by Anthropic. None of it has been independently verified by a third party. The company also says these are a hand picked selection of the most sophisticated cases it found, not a representative sample. Keep both facts in mind as you read.&lt;/p&gt;
&lt;h2 id=&quot;what-the-report-actually-covers&quot;&gt;What the report actually covers&lt;/h2&gt;
&lt;p&gt;The report, &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.anthropic.com/threat-intelligence-report-september-2026&quot;&gt;Detecting and Countering Misuse of AI, September 2026&lt;/a&gt;, covers activity Anthropic says it detected and disrupted between December 2025 and August 2026, sorted into 7 categories, which are cyber operations, influence operations, surveillance, scams and fraud, biological misuse, conventional weapons, and illicit distillation. Claude Haiku, Sonnet, and Opus are the models named in most cases.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-01.Dg1Iy2cD_ZXm3Yh.webp&quot; alt=&quot;Figure 3. The 7 harm categories the report is organized around.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1185&quot; height=&quot;604&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Figure 3. The 7 harm categories the report is organized around. Source: Anthropic, Detecting and Countering Misuse of AI, Sep 2026&lt;/em&gt;&lt;/p&gt;
&lt;p&gt;The pattern Anthropic keeps returning to is that the techniques here are old, and the economics are new. Nothing in the worst cases required knowledge that did not already exist somewhere. What changed is that a single person, or a tiny cell, got the rough equivalent of a full engineering team for the price of an API subscription.&lt;/p&gt;
&lt;p&gt;The scale underneath those categories is worth stating plainly. Anthropic uses roughly 40 internal case designators, which it calls Generative Threat Groups, across the report. It documents 9 influence operations spread over 6 continents, a dating app network running more than 4,700 AI personas in romantic conversations with at least 25,000 real people, 6 conventional weapons cases, and 5 biological ones. One detail I keep thinking about is which models show up. Every case involved Claude Haiku, Sonnet, or Opus. Anthropic’s newer Fable and Mythos class models appear in exactly 1 incident in the entire document, which the company credits to the stronger safeguards on those models. This is Anthropic’s 4th threat report, after ones in March, August, and November 2025, and a February 2026 distillation specific disclosure that named 3 labs and counted more than 16 million exchanges.&lt;/p&gt;
&lt;h2 id=&quot;the-cyberattacks-now-run-themselves-mostly&quot;&gt;The cyberattacks now run themselves, mostly&lt;/h2&gt;
&lt;p&gt;The theme Anthropic hammers hardest is that AI stopped being an adviser and became the thing executing the attack. In most of the cyber cases, operators pointed a multi agent setup at a general goal and let it handle reconnaissance, exploitation, credential theft, and data exfiltration, with humans stepping in mainly to pick targets and check results.&lt;/p&gt;
&lt;p&gt;The clearest example is GTG-20006, an actor whose tradecraft Anthropic says is consistent with Midnight Blizzard, Microsoft’s name for the Russian intelligence linked group behind the 2020 SolarWinds breach. Anthropic is careful with that phrase, since “consistent with” stops deliberately short of naming them outright. Over about 130 days, the operation engaged 24 of 27 targeted institutions, including Ukrainian ministries, defense bodies, and drone supply chain manufacturers. The malware monitored itself and recompiled its own code when security tools flagged it. Anthropic says the operation also touched hotel guest WiFi networks and moved more than 300,000 North African national ID records out.&lt;/p&gt;
&lt;p&gt;Then there is the criminal side. Anthropic ties one extortion cluster to affiliates it associates with ShinyHunters, who completed cloud compromises in as little as 2 to 3 hours. One operator spun up 10 AWS compute workers to decompile 1.8 million Android apps looking for embedded secrets. Another pulled more than 2,100 Azure AD tokens across over 40 corporate tenants in 34 hours. A separate China based group automated firmware analysis on security appliances and turned up more than a dozen potential zero day vulnerabilities in one month across roughly 50 targets. The report’s phrase for the general style is “vibe hacking,” where an operator hands the model a goal and lets it improvise the steps.&lt;/p&gt;
&lt;h2 id=&quot;chinese-labs-ran-nearly-200-million-exchanges-through-claude&quot;&gt;Chinese labs ran nearly 200 million exchanges through Claude&lt;/h2&gt;
&lt;p&gt;The distillation section is the one that drew the most coverage, and the numbers are large. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.cnbc.com/2026/09/11/chinese-ai-labs-moonshot-deepseek-alibaba-anthropic.html&quot;&gt;Anthropic says&lt;/a&gt; 7 China based AI labs ran unauthorized campaigns to extract Claude’s capabilities for training their own models, generating roughly 190 million exchanges in total.&lt;/p&gt;
&lt;p&gt;Alibaba accounts for the bulk of that. Anthropic says accounts tied to Alibaba ran more than 151 million exchanges with Claude between May and July 2026, peaking near 3 million a day across more than 3,500 accounts it flagged as fraudulent. The company calls it the largest distillation campaign it has ever measured. The method was specific. A fixed prompt injected into every request forced Claude Opus 4.6 and 4.7 to write out their full chain of thought reasoning inside inline tags, and the pipeline scraped only the text inside those tags. Those reasoning traces went into fine tuning data for Qwen 3.5, 3.6, and 3.7. Anthropic says Alibaba also used Claude for its own research, including building reinforcement learning environments and model architecture work. Using Claude to help build the thing that competes with Claude.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-02.C5mqBX4W_2azVA.webp&quot; alt=&quot;Figure 1. Exchanges Anthropic attributes to each lab. Alibaba dwarfs the rest.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1185&quot; height=&quot;668&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Figure 1. Exchanges Anthropic attributes to each lab. Alibaba dwarfs the rest. Source: Anthropic, Sep 10 2026 (alleged, not independently verified)&lt;/em&gt;&lt;/p&gt;
&lt;p&gt;The Moonshot case is the one that made me sit up. Anthropic says Moonshot silently forwarded around 300,000 customer requests to Claude over a 10 day window through thousands of fake accounts, mostly to Opus, and showed users Claude’s answers as if Kimi had produced them. DeepSeek did something similar but more targeted, screening inbound requests for signals that a user was coming through a coding harness like Claude Code, then routing those users to Claude Opus. Anthropic says it counted more than 12.1 million exchanges tied to DeepSeek over 14 days in July.&lt;/p&gt;
&lt;p&gt;Sit with what that means if you were a Kimi or DeepSeek customer in that window. You typed something into a Chinese AI product, and Anthropic says it was quietly handed to an American company you never agreed to send anything to. Some of those exchanges, Anthropic says, carried sensitive data belonging to individuals, multinational companies, and state affiliated organizations. Among the rerouted DeepSeek traffic, the report describes one engineer’s specification for a municipal Public Security Bureau case management system.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-03.DP4Gb-Z-_Zqoz11.webp&quot; alt=&quot;Figure 2. Users Anthropic says were routed to Claude without being told.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1110&quot; height=&quot;603&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Figure 2. Users Anthropic says were routed to Claude without being told. Source: Anthropic, Sep 10 2026&lt;/em&gt;&lt;/p&gt;
&lt;p&gt;4 more labs get named. Anthropic says Xiaomi replayed more than 400,000 of its own users’ coding sessions through Claude to make training data. Zhipu, known internationally as &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;http://z.ai/&quot;&gt;Z.ai&lt;/a&gt;, rotated through 273 accounts and pushed more than 770,000 exchanges through an automated reasoning cleaner tool to train its GLM-5.3 model, and interestingly went after Claude Fable first, gave up when the safeguards degraded the output, then switched to an older Claude model and another provider. SenseTime, by Anthropic’s account, skipped the work and simply bought harvested transcripts from third party data brokers. The seventh, MiniMax, ran an unbranded proxy service through a shell company offering only US models, which Anthropic assessed was likely built to collect user exchanges for training rather than proving that intent outright.&lt;/p&gt;
&lt;p&gt;China &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.rappler.com/technology/anthropic-threat-intelligence-report-september-2026/&quot;&gt;rejected all of it&lt;/a&gt;. The Commerce Ministry said the allegations had no factual or legal basis and warned that countermeasures could follow, and the foreign ministry accused Anthropic of distorting facts. Alibaba later prohibited its employees from using Claude Code. Beijing’s other line was that distillation is a neutral technique, and that US companies have distilled Chinese models too.&lt;/p&gt;
&lt;p&gt;That last point deserves a fair hearing. Distillation is a normal training method. What Anthropic describes is the covert, industrial scale version run through fake accounts and stolen cards, which is a different animal. I think that distinction is real, and I also think it is going to get flattened in every retelling until “distillation” and “theft” read as the same word.&lt;/p&gt;
&lt;h2 id=&quot;a-weapons-cell-in-yemen-used-claude-as-its-engineering-team&quot;&gt;A weapons cell in Yemen used Claude as its engineering team&lt;/h2&gt;
&lt;p&gt;This is the case that stays with you. Anthropic says it &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.bloomberg.com/news/articles/2026-09-11/anthropic-says-yemen-group-used-claude-in-missile-development&quot;&gt;identified a cell in northern Yemen&lt;/a&gt;, tracked under the designator GTG-87001, running 3 weapons programs at once, a guided rocket using a commercially available phone class flight computer, a multi stage ballistic missile with a stated range goal over 2,000 km, and a missile family it calls R2000 that included a hypersonic glide vehicle variant.&lt;/p&gt;
&lt;p&gt;Anthropic did not name the Houthis. The cell operates in Houthi controlled territory, and later reporting from Bloomberg and others has drawn the link, but I want to be precise about what the report claims and does not claim. Anthropic says the actors used Claude Code for guidance, navigation, and control software, in its own words “in place of human software engineers,” running several instances in parallel with different roles and splitting the work across sessions so no single prompt revealed the whole picture. Then they test fired the guided rocket. It failed. Within hours they were back in Claude asking why.&lt;/p&gt;
&lt;p&gt;I laughed at that, then felt bad about laughing. A group builds a rocket with a chatbot, the rocket does not work, and they open a fresh session to debug it like a broken deployment script.&lt;/p&gt;
&lt;p&gt;The scary headlines leave out the part that matters most. Anthropic &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://thedefensewatch.com/middle-east-defense-security/claude-code-yemen-missile-development/&quot;&gt;explicitly said it had no evidence&lt;/a&gt; the cell ever fielded an operational weapon. The only physical test in the report is a guided rocket that appears to have failed. Anthropic did not formally establish that the Houthis controlled the cell. What the report does establish is narrower. A weapons development cell existed in northern Yemen, it used Claude for guidance and simulation work, it ran one failed live test, it tried to evade safeguards, and it had already built an offline simulation toolkit that runs without Claude or any commercial tool like MATLAB. That toolkit persists whether or not the accounts get banned, which is the genuinely uncomfortable detail, more than the missile ranges.&lt;/p&gt;
&lt;p&gt;The conventional weapons section has more. Anthropic describes likely freelance Russian actors building an autonomous FPV drone swarm with a small onboard model that could select human targets and detonate without a person in the loop, trained on captured Ukrainian combat footage. It describes a Chinese actor building roughly 16 modules for electronic warfare and air defense suppression, where the default simulation scenario at one point switched to 12 targets positioned in Taiwan.&lt;/p&gt;
&lt;h2 id=&quot;a-surveillance-state-built-by-one-consultant&quot;&gt;A surveillance state, built by one consultant&lt;/h2&gt;
&lt;p&gt;The surveillance section is quieter than the missile one, and in some ways it unsettled me more. Anthropic describes a single consultant who used Claude to help build a platform called Lakana 360 for Mali’s state intelligence service. As designed, it would cover roughly 25 million SIM cards across all 3 of the country’s mobile operators. The feature list is the part that lands. Warrant free dossier generation. Voice identification that recognizes the same speaker across different phone numbers. Detection of anyone using a VPN or encryption. Geofenced watchlists that alert an analyst when a flagged person enters an area. Integration with the national identity registry. One person, with a coding model, drafting the machinery of a national watch apparatus.&lt;/p&gt;
&lt;p&gt;Anthropic lists more in this category. It says China based security bureaus generated 2,475 investigative briefs in 30 days covering dissidents, Uyghurs, pro democracy protests abroad, and religious figures, and that one bureau used Claude to draft an internal manual on using AI for surveillance. It describes Iranian units analyzing 155,216 tweets and producing open source intelligence on US naval fleet positions, and a commercial vendor profiling diaspora populations with synthetic accounts. The common thread is the one Anthropic keeps returning to, that a single operator or a small team now produces the software, interfaces, and analysis pipelines that used to need a whole staff.&lt;/p&gt;
&lt;h2 id=&quot;did-anthropic-read-peoples-chats-to-write-this&quot;&gt;Did Anthropic read people’s chats to write this?&lt;/h2&gt;
&lt;p&gt;This was the top question in the replies. The report is built from investigations of accounts that were already flagged. Detection starts with automated classifiers, metadata, and abuse signals like proxy networks, unsupported regions, extraction patterns, and safety flags. When those fire, investigators look at the associated prompts and sessions to attribute the activity, ban the accounts, and describe what happened.&lt;/p&gt;
&lt;p&gt;So did Anthropic read user chats? Yes, the chats of investigated and banned accounts. Did Anthropic read everyone’s chats to write this report? No. Its stated default is that staff cannot browse conversations unless something is flagged, legally required, or routed through a controlled review path. The report is evidence of targeted safety investigation, not routine human reading of every Claude conversation.&lt;/p&gt;
&lt;p&gt;I do not think the discomfort is irrational, though. The detail here goes down to Telegram bot IDs and individual operator aliases, and Anthropic’s recent record on privacy does not fully settle the nerves. The company &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.aljazeera.com/news/2026/9/11/anthropic-claims-claude-ai-used-for-missile-projects-global-espionage&quot;&gt;called San Francisco police&lt;/a&gt; on a user over a threatening chat and then reportedly declined to hand over the messages. Its privacy policy was updated to make sharing conversation data with law enforcement easier at its own discretion. In July, a flaw exposed shared Claude conversations to search engines. If you are a developer, a researcher, or someone asking Claude about a health or money problem, “we only look when something is flagged” still requires trusting whoever writes the flagging rules.&lt;/p&gt;
&lt;h2 id=&quot;not-everyone-thinks-publishing-this-was-a-public-service&quot;&gt;Not everyone thinks publishing this was a public service&lt;/h2&gt;
&lt;p&gt;The criticism splits into 2 lines, and I find both worth taking seriously.&lt;/p&gt;
&lt;p&gt;The first is that the report is fear marketing. Everything in it came from publicly available information, the argument goes, and bad actors do bad things with or without a chatbot. The report’s own defense is the economics point I mentioned, that the techniques are old and the cost collapsed, which I think is the stronger side of this particular argument. Nothing in the Yemen case needed knowledge that did not already exist. What changed is that a small cell got the equivalent of a software team on tap.&lt;/p&gt;
&lt;p&gt;The second criticism is about the messenger, and it is harder to wave off. Anthropic is a commercial company publishing unverified findings that happen to position it as the responsible adult in the room, while it competes directly with several of the labs it is naming and lobbies Washington on export policy. Alibaba’s US listed shares dropped toward a 52 week low after an earlier version of these allegations circulated in June. A report can be accurate and self serving at the same time, and this one reads as both. When every disclosure lands as “everyone is attacking us, everyone is copying us, everyone needs us,” a reasonable person is allowed to raise an eyebrow.&lt;/p&gt;
&lt;p&gt;2 more findings complicate the tidy China versus America framing. Research &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.ctgt.ai/research&quot;&gt;published in July by CTGT&lt;/a&gt; found that censorship baked into Chinese models does not reliably survive distillation into models trained on them, which undercuts one of Washington’s main arguments against Chinese open weights. A separate &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://fortune.com/2026/08/ai-models-political-censorship-study/&quot;&gt;study covered by Fortune&lt;/a&gt; looked at 10 models and found refusal rates for political criticism averaging 34% for models from restrictive countries and 14% for models from freer ones, with wide variation between individual models. American models are not neutral ground either.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-04.DgK7s699_ZyjcP3.webp&quot; alt=&quot;Figure 4. Refusal rates for political criticism, restrictive versus freer countries.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1035&quot; height=&quot;603&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Figure 4. Refusal rates for political criticism, restrictive versus freer countries. Source: Study covered by Fortune, 2026&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;why-this-is-really-an-argument-about-rules&quot;&gt;Why this is really an argument about rules&lt;/h2&gt;
&lt;p&gt;I keep coming back to one fact about this report. Anthropic caught these cases, and I believe it caught the ones it describes. But every single case is one where a commercial company decided, on its own timeline and by its own rules, what counted as misuse and what to do about it. No external body saw any of this before it went live.&lt;/p&gt;
&lt;p&gt;Nothing in these 154 pages involves runaway self improving AI. It is a cell in Yemen with a chatbot, a hacktivist with stolen API keys running an operation that used to need a team, and a drone that picks its own target. That is the near term version of the control problem, and it is already here, not in 2030.&lt;/p&gt;
&lt;p&gt;There are proposals floating for the longer term. Sam Altman floated a compute cap for frontier development, which has the same shape of weakness as every arms control idea, since it lowers risk only if the major powers, China included, accept verifiable rules, and neither the participation nor the verification exists right now.&lt;/p&gt;
&lt;p&gt;I think the takeaway is simple to state and hard to do. The big AI companies need real constraints on their own models, and governments need rules that do not depend on a lab volunteering the information when it suits them. Regulation will lag the technology no matter what. The gap only grows while we wait, so the sensible time to start closing it was before a cell in Yemen filed a bug report against a rocket.&lt;/p&gt;</content:encoded><dc:creator>Shivashish Yadav</dc:creator><category>Security</category><category>anthropic-claude</category><category>claude</category><category>alibaba</category><category>deepseek</category><author>hello@soursen.org (Shivashish Yadav)</author></item><item><title>An AI Read Zcash’s Code for One Day. Then It Found the Bug That 4 Years of Human Audits Missed.</title><link>https://soursen.org/articles/an-ai-read-zcashs-code-for-one-day-then-it-found-the-bug-that-4-years/</link><guid isPermaLink="true">https://soursen.org/articles/an-ai-read-zcashs-code-for-one-day-then-it-found-the-bug-that-4-years/</guid><description>Here is a thought experiment worth sitting with. Imagine a piece of software has been audited repeatedly by some of the best cryptographers in the world. It has been live in production for 4 years. Hundreds of millions of dollars flow through it.</description><pubDate>Mon, 15 Jun 2026 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;Here is a thought experiment worth sitting with. Imagine a piece of software has been audited repeatedly by some of the best cryptographers in the world. It has been live in production for 4 years. Hundreds of millions of dollars flow through it. The people who designed it, built it, reviewed it, and staked their professional reputations on it have all signed off.&lt;/p&gt;
&lt;p&gt;Now imagine that an AI model reads that same code on the day it is publicly released, roughly 24 hours after it becomes available to anyone with an internet connection, and identifies a flaw so significant that it could have allowed an attacker to silently mint unlimited counterfeit currency for the entire four years the system was running.&lt;/p&gt;
&lt;p&gt;That is not a thought experiment. That is what happened to Zcash in late May 2026. The implications reach well beyond one privacy coin, because what it tells us about the gap between human and machine pattern recognition in security-critical code is genuinely unsettling.&lt;/p&gt;
&lt;h2 id=&quot;what-zcash-is-and-why-this-matters-more-than-a-typical-bug&quot;&gt;What Zcash Is, and Why This Matters More Than a Typical Bug&lt;/h2&gt;
&lt;p&gt;&lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://z.cash/&quot;&gt;Zcash&lt;/a&gt; is a cryptocurrency built around one central promise: complete financial privacy.&lt;/p&gt;
&lt;p&gt;While Bitcoin records every transaction on a fully public ledger where anyone can trace the movement of funds, Zcash offers what it calls shielded transactions, where the sender, recipient, and amount are all hidden. The technology underpinning this is called a zero-knowledge proof, a mathematical construction that lets one party prove to another that a statement is true without revealing any of the underlying information required to verify it.&lt;/p&gt;
&lt;p&gt;In practical terms, a zero-knowledge proof allows Zcash to say this transaction is valid, the coins being spent exist, the arithmetic balances, without disclosing any of the specifics to anyone. It is elegant, mathematically powerful, and widely regarded as one of the more sophisticated privacy systems in production use. The specific implementation Zcash uses is called the &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://zips.z.cash/zip-0224&quot;&gt;Orchard Shielded Pool&lt;/a&gt;, which launched in May 2022 as the network’s most advanced privacy layer.&lt;/p&gt;
&lt;p&gt;The entire system rests on a property cryptographers call soundness. Soundness is the guarantee that you cannot construct a valid proof for a false statement. If the proof says a transaction is legitimate, the transaction must actually be legitimate. That guarantee is not a nice-to-have. It is the mathematical foundation the entire privacy model sits on. Without it, the system does not just have a problem. It has no reliable basis.&lt;/p&gt;
&lt;p&gt;The bug that sat undetected for 4 years was a crack in exactly that foundation.&lt;/p&gt;
&lt;h2 id=&quot;two-lines-of-code-4-years-and-unlimited-counterfeit-currency&quot;&gt;Two Lines of Code, 4 Years, and Unlimited Counterfeit Currency&lt;/h2&gt;
&lt;p&gt;The vulnerability lived inside a Rust library called &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://github.com/zcash/halo2&quot;&gt;halo2&lt;/a&gt;, specifically in a component called the elliptic curve multiplication gadget. Elliptic curve cryptography is the mathematical backbone of most modern blockchain systems, including Bitcoin and Ethereum, used to verify ownership and authenticate transactions. In Zcash’s zero-knowledge circuit, this gadget performs a specific calculation as part of proving that a shielded transaction is valid.&lt;/p&gt;
&lt;p&gt;The problem was that the gadget failed to properly constrain its inputs. In plain language, the circuit was supposed to accept only mathematically valid inputs. It was not doing that. Inputs that should have been rejected were passing the check. Cryptographers call this being under-constrained, and in a zero-knowledge proof system, being under-constrained is not a minor edge case. It means the soundness guarantee has been broken.&lt;/p&gt;
&lt;p&gt;The consequence of that broken guarantee was severe. An attacker who understood the flaw could construct proofs for transactions that were mathematically invalid but would still pass verification. They could create Zcash inside the shielded pool that did not correspond to any real coins. Counterfeit currency, minted in unlimited quantities, completely invisible on the public ledger because the privacy mechanism that hides legitimate transactions would hide the fake ones equally well.&lt;/p&gt;
&lt;p&gt;A real coin and a fake coin would be indistinguishable. Not just to outside observers. To anyone, including the Zcash developers themselves.&lt;/p&gt;
&lt;p&gt;This vulnerability was present from the day the Orchard Shielded Pool launched in May 2022. It survived through multiple rounds of professional security review. The people who audited Zcash during this period were not amateurs. The codebase had been examined by cryptographers from firms like &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.trailofbits.com/&quot;&gt;Trail of Bits&lt;/a&gt;, one of the most respected names in blockchain security auditing. Two lines of code. 4 years. Every human reviewer missed it.&lt;/p&gt;
&lt;h2 id=&quot;how-an-ai-found-in-hours-what-experts-missed-in-years&quot;&gt;How an AI Found in Hours What Experts Missed in Years&lt;/h2&gt;
&lt;p&gt;Taylor Hornby is a security researcher. In April 2026, he was hired by &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://shieldedlabs.net/&quot;&gt;Shielded Labs&lt;/a&gt;, an organisation focused on Zcash’s long-term development, specifically to audit Zcash’s cryptographic infrastructure. His brief was to look for problems proactively rather than wait for one to surface.&lt;/p&gt;
&lt;p&gt;&lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.anthropic.com/news/claude-opus-4-8&quot;&gt;Anthropic released Claude Opus 4.8&lt;/a&gt; on May 28, 2026. Hornby discovered the bug on May 29. The gap between those two events is approximately 24 hours.&lt;/p&gt;
&lt;p&gt;The important detail here is how Hornby used the model, because he did not simply paste the code into a chat window and ask what was wrong. He built what he described as a custom auditing framework, reportedly calling it a Zcash full-stack auditor, with carefully constructed prompts designed specifically to hunt for constraint failures in zero-knowledge circuits. He was directing the AI toward a specific class of vulnerability, using its pattern recognition capabilities within a structured methodology he had designed for the purpose.&lt;/p&gt;
&lt;p&gt;The AI identified the under-constrained elliptic curve multiplication gadget. It then went further, helping Hornby write a working proof-of-concept exploit. When he ran it in a local test environment, the exploit succeeded. It generated unlimited counterfeit Zcash, invisible on-chain, exactly as the theoretical analysis predicted it would.&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Shielded Labs confirmed the exploit worked. They said they believed Hornby had found it before any attacker did. The word they used was probably. That word is doing a lot of work.&lt;/em&gt;&lt;/p&gt;
&lt;p&gt;This raises a question that extends beyond Zcash. If a model that had been publicly available for one day could identify a vulnerability of this magnitude in one of the most scrutinised codebases in crypto, the security assumptions underlying every comparable system become uncertain. Hornby’s methodology, combining structured prompting with AI pattern recognition to specifically target constraint failures, is reproducible. The security research community is now working through the implications. Reports suggest Hornby may be turning his attention to other privacy coin codebases, including &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.getmonero.org/&quot;&gt;Monero&lt;/a&gt;, though no formal announcement has been confirmed.&lt;/p&gt;
&lt;h2 id=&quot;the-response-fast-competent-and-still-not-enough-to-resolve-the-core-problem&quot;&gt;The Response: Fast, Competent, and Still Not Enough to Resolve the Core Problem&lt;/h2&gt;
&lt;p&gt;Credit where it is due. Once the vulnerability was confirmed, the Zcash development team moved with unusual speed.&lt;/p&gt;
&lt;p&gt;On June 1 and 2, developers deployed an emergency soft fork that temporarily disabled all Orchard transactions, closing the attack window while a permanent fix was prepared. On June 3, the &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://zips.z.cash/zip-0253&quot;&gt;NU6.2 hard fork&lt;/a&gt; activated with a corrected circuit. The reason a hard fork was required, rather than a simpler update, is instructive. Changing even two lines of a zero-knowledge circuit changes its cryptographic verifying key. Every node in the network must upgrade to the new key to verify transactions correctly. There is no way to patch this class of vulnerability without requiring the entire network to move simultaneously.&lt;/p&gt;
&lt;p&gt;The transition was not seamless. Block explorer synchronisation failures lasted over 4 hours, creating the visual appearance of a network halt, though the Zcash mainnet itself continued producing blocks throughout. Orchard transactions remained suspended for approximately 24 hours across the full remediation window.&lt;/p&gt;
&lt;p&gt;When the dust settled, the official position was that no evidence of exploitation had been found. That statement is true. It is also, from a cryptographic standpoint, essentially the only statement that could ever be made, regardless of what actually happened.&lt;/p&gt;
&lt;h2 id=&quot;the-paradox-at-the-centre-of-every-privacy-coin&quot;&gt;The Paradox at the Centre of Every Privacy Coin&lt;/h2&gt;
&lt;p&gt;This is the part of the story that does not get resolved by a fast patch or a competent development team, because it is not a bug. It is a structural feature of the design.&lt;/p&gt;
&lt;p&gt;Orchard’s privacy protections hide all transaction amounts and all participant identities from everyone, including the developers. When the Zcash Foundation said they found no evidence of exploitation, what they meant precisely is that they found no public-chain signals that exploitation occurred. They cannot say more than that, because the privacy mechanism that makes Zcash valuable makes that kind of verification impossible.&lt;/p&gt;
&lt;p&gt;Consider what that means in practice. If an attacker had identified this vulnerability independently before Hornby, minted a large quantity of counterfeit Zcash inside the shielded pool, and simply held those coins without moving them, the turnstile mechanism that tracks value crossing between pools would show nothing unusual. The attacker would be sitting on an invisible reserve of currency that should not exist, with no verifiable way for anyone to detect it.&lt;/p&gt;
&lt;p&gt;The Zcash developers do point to the turnstile as a meaningful safeguard. If an attacker tried to profit by moving counterfeit coins to a transparent address or an exchange, the turnstile would register a surplus that does not reconcile with the legitimate supply. That is a real defence. Its limitation is that it only functions if the attacker moves the coins. A patient attacker who never moves them remains invisible indefinitely.&lt;/p&gt;
&lt;blockquote&gt;
&lt;p&gt;The exact feature that protects your financial privacy is the same feature that makes it impossible to verify the total supply of the currency you are holding. Those two properties are not in tension. They are the same property, viewed from two different angles.&lt;/p&gt;
&lt;/blockquote&gt;
&lt;p&gt;This is not specific to Zcash. It is the fundamental trade-off of any fully private monetary system. Bitcoin’s entire value proposition as sound money rests partly on the fact that you can verify the supply at any time. Anyone can check the blockchain and confirm that exactly 21 million Bitcoin will ever exist. With a fully shielded system, that kind of verification is by design impossible. You are trusting a mathematical proof system to be correct. This incident demonstrated that those proof systems can contain flaws that survive years of expert review.&lt;/p&gt;
&lt;h2 id=&quot;what-the-market-said-and-what-that-tells-you&quot;&gt;What the Market Said, and What That Tells You&lt;/h2&gt;
&lt;p&gt;Before the full disclosure on approximately June 5, &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://coinmarketcap.com/currencies/zcash/&quot;&gt;ZEC was trading between $620 and $640&lt;/a&gt;. Within a day or two of the vulnerability becoming public, the price had fallen to intraday lows between $255 and $310. That is a decline of roughly 50% to 57%. Approximately $5 billion in market capitalisation was erased.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-01.btfF7UhX_Z1DreFe.webp&quot; alt=&quot;Illustration for the section “What the Market Said, and What That Tells You”&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;744&quot; height=&quot;269&quot;&gt;&lt;/p&gt;
&lt;p&gt;What analysts noted about the sell-off was its character. This was not the leverage-driven cascade that typically characterises large crypto price drops, where futures positions get liquidated in sequence, amplifying the move. The selling was primarily spot-driven, meaning actual holders selling actual coins, which indicates genuine loss of conviction rather than a technical market event. Liquidations were substantial, with some trackers reporting nine-figure totals in a single twenty-four-hour window, though exact figures were not independently confirmed.&lt;/p&gt;
&lt;p&gt;&lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://cypherpunk.holdings/&quot;&gt;Cypherpunk Holdings&lt;/a&gt;, a Winklevoss-backed company listed on NASDAQ that holds Zcash as a treasury asset, watched its stock fall approximately 37% on disclosure, closing at $59. The company held approximately 303,900 ZEC at an average cost basis of roughly $333 per coin, putting the position briefly and deeply underwater.&lt;/p&gt;
&lt;h2 id=&quot;the-people-who-sold&quot;&gt;The People Who Sold&lt;/h2&gt;
&lt;p&gt;Arthur Hayes, co-founder of BitMEX and a well-known figure in crypto markets, &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://twitter.com/CryptoHayes&quot;&gt;disclosed publicly&lt;/a&gt; that he liquidated his entire Zcash position on the day of disclosure. His reasoning was specific and worth examining on its own terms. He was not claiming an exploit had definitely occurred. He was arguing that the privacy thesis for a coin like Zcash depends on a narrative of mathematical perfection, not mathematical probability. In his framing, a supply that cannot be cryptographically verified is a broken promise to holders, regardless of how unlikely actual exploitation might be.&lt;/p&gt;
&lt;p&gt;On-chain researcher ZachXBT responded by pointing out that Hayes had publicly exited four previously hyped positions within roughly fifteen days, raising the question of whether followers were being used as exit liquidity. The criticism landed. Hayes’s standing in crypto circles took a visible hit in the days following.&lt;/p&gt;
&lt;h2 id=&quot;the-people-who-held&quot;&gt;The People Who Held&lt;/h2&gt;
&lt;p&gt;Cypherpunk Technologies and Cameron Winklevoss reaffirmed their position. Their argument was that the bug had been found by a researcher the project itself commissioned, not by a hostile attacker, and that the patch was deployed within days of discovery. They characterised this as evidence of a mature security culture rather than a sign of fundamental failure.&lt;/p&gt;
&lt;p&gt;Grayscale’s legal team made a statistical argument: any attacker sophisticated enough to have independently found and exploited this vulnerability would almost certainly have cashed out during the significant Zcash bull run that occurred prior to disclosure, rather than sitting silently on an invisible reserve. The counterargument is that we are assuming rationality and impatience on the part of a hypothetical attacker. Neither assumption can be verified. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://grayscale.com/insights/zcash-security-assessment-2026&quot;&gt;Grayscale’s public statement&lt;/a&gt; laid out their reasoning in detail.&lt;/p&gt;
&lt;p&gt;Both positions are defensible. That is precisely what makes the situation uncomfortable. The honest answer is that nobody knows, not in the probabilistic sense of unlikely, but in the stronger sense of unknowable with the information available.&lt;/p&gt;
&lt;h2 id=&quot;what-comes-next-and-what-can-never-be-undone&quot;&gt;What Comes Next, and What Can Never Be Undone&lt;/h2&gt;
&lt;p&gt;The Zcash development team has announced an upgrade called Ironwood, targeted for late July 2026, which includes a strict turnstile accounting mechanism. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://zips.z.cash/zip-0255&quot;&gt;The Ironwood proposal&lt;/a&gt; would force all coins migrating from the old Orchard pool through a publicly auditable checkpoint, enabling what amounts to a verifiable census of the supply at that moment.&lt;/p&gt;
&lt;p&gt;Understanding precisely what Ironwood can and cannot establish is important. If the supply reconciles cleanly when coins migrate through the new checkpoint, that is the strongest possible signal that the books are honest. It is not cryptographic proof that exploitation did not occur, because it cannot account for coins that were minted and are being held inside the old pool without ever moving. It is, however, evidence that is currently absent, and its presence or absence after the migration will be informative.&lt;/p&gt;
&lt;p&gt;What Ironwood cannot do is retroactively verify the history of the Orchard pool from May 2022 to the present. That history is sealed by the same privacy guarantees that sealed it before the bug was found. The window of uncertainty covers four years. It does not close when Ironwood activates.&lt;/p&gt;
&lt;p&gt;The broader question the security research community is now sitting with is less about Zcash specifically and more about what Hornby’s methodology implies for other systems. Zero-knowledge proof circuits are complex, difficult to audit, and increasingly deployed in high-value contexts. The &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://ethereum.org/en/developers/docs/scaling/zk-rollups/&quot;&gt;Ethereum Foundation has invested heavily in ZK-rollup technology&lt;/a&gt;, which uses related cryptographic constructions. If constraint failures of this kind can survive four years of expert review in one of the most watched codebases in privacy-focused crypto, the audit confidence levels assigned to comparable systems deserve reconsideration.&lt;/p&gt;
&lt;h2 id=&quot;the-question-this-actually-raises&quot;&gt;The Question This Actually Raises&lt;/h2&gt;
&lt;p&gt;What the Zcash incident is really asking you to think about is not whether Zcash is safe to hold or whether an exploit happened. It is a more fundamental question about what you are trusting when you hold a fully private asset.&lt;/p&gt;
&lt;p&gt;With Bitcoin or Ethereum, the trust model is explicit and auditable. The supply is public. The transaction history is public. Anyone with the technical capability can verify that the rules of the system are being followed. The transparency is not incidental. It is the mechanism by which trust is established without requiring you to trust any particular party.&lt;/p&gt;
&lt;p&gt;A fully shielded privacy coin inverts that model. The privacy is the product. The opacity is the feature. You are trusting the mathematics, the implementation of the mathematics, and the audit history of that implementation to all be correct. This incident did not break the mathematics. It found a flaw in the implementation of the mathematics, a flaw that survived the audit history. The mathematics itself, the underlying zero-knowledge proof theory, remains sound.&lt;/p&gt;
&lt;p&gt;The question is whether that distinction matters to you as a holder. A system can be theoretically sound and still contain implementation errors that undermine its guarantees in practice. That is true of every complex software system ever built. What is unusual here is that the privacy mechanism makes it impossible to determine whether any particular error was exploited.&lt;/p&gt;
&lt;p&gt;The Ironwood upgrade in late July will provide the clearest available signal. If the supply reconciles, the most likely interpretation is that the books are honest. If it does not reconcile cleanly, the implications are severe and probably irreversible. Either way, the four years before that migration are not something any audit can retroactively resolve.&lt;/p&gt;
&lt;p&gt;What Hornby and his AI-assisted methodology demonstrated is that the tools available for finding these vulnerabilities have changed substantially. The question that follows is whether the tools available for verifying that they were never exploited have changed at all. The answer, in a fully private system, is that they have not, because the privacy that needs protecting and the transparency that enables verification are in direct conflict by design.&lt;/p&gt;
&lt;p&gt;That is not a solvable problem. It is the trade-off you accept when you hold a privacy coin. This incident just made that trade-off harder to ignore.&lt;/p&gt;</content:encoded><dc:creator>Shivashish Yadav</dc:creator><category>Security</category><category>ai</category><category>cryptocurrency</category><category>artificial-intelligence</category><category>web3</category><category>hacking</category><author>hello@soursen.org (Shivashish Yadav)</author></item></channel></rss>