Unreleased Claude model built a one-click Chrome exploit

An unreleased Anthropic research model found the bugs behind 62 Chrome CVEs and built five working attack chains, among them one-click code execution in Chrome and zero-click root on Linux servers, Anthropic said in an October 27 post. If your team manages Chrome, the build to look for is 156.0.8078.73, which shipped the same day with 18 more Anthropic-credited fixes, four of them rated critical. The Safari, Android, Windows and Linux bugs are already fixed, the Windows one since April.
The other three chains targeted Safari, Android and Windows, covering open-source and proprietary code alike, and each needed little or no user interaction. Anthropic reported every bug under its coordinated disclosure policy. Vendors get 90 days to patch, then another 45 before Anthropic discusses details beyond the public advisory.
That policy explains the timing. In April, when Anthropic introduced Mythos Preview alongside Project Glasswing, a critical-software security effort announced April 7 whose partners include Google, Apple, Microsoft and the Linux Foundation, it said the model had already found high-severity bugs in every major operating system and web browser, The Hacker News reported. The Frontier Red Team’s April write-up held back browser details because nothing had been patched yet. This post is the first to give them.
What each chain asks of the victim
The Chrome chain needs one click. On a stock macOS Chrome, the proof of concept opens Calculator outside the browser sandbox within seconds of the page loading. It strings four bugs together: heap corruption in V8, Chrome’s JavaScript engine; a second V8 bug that breaks out of V8’s memory sandbox and defeats ASLR; a race condition in CacheStorage that leaks the memory layout of the privileged browser process; and a use-after-free in FedCM, Chrome’s federated sign-in feature, which runs the attacker’s code outside the sandbox. Anthropic counts 13 sandbox escapes and 21 renderer-compromising bugs among the 62 CVEs.
Safari’s chain never runs attacker code at all. A use-after-free in JavaScriptCore, combined with a way around WebKit’s Gigacage isolation, gives it read and write access to browser memory. It then rewrites a security flag, an object size and a URL to switch off the same-origin policy, the browser rule that keeps one site from reading another site’s data. After that, any page could read a signed-in user’s email or bank balance and send authenticated requests on their behalf. Anthropic says several of the weaknesses are intended behaviors that only become exploitable in combination. Apple acknowledged the report the day it arrived.
The Linux exploit needs no interaction. A remote attacker sends packets to a server with a multipath TCP listener and runs commands as root, without credentials. Losing the race only fails the handshake and doesn’t crash the system, so an attacker can retry indefinitely. In Anthropic’s proof of concept, leaking a kernel pointer took 48 minutes, and reaching command execution took five more.
The other two chains are narrower. The Android chain pulls photos and personal data over USB from a stolen, locked Pixel 9 by combining three bugs across open-source Android and Google Play services, including one that skips the six-digit pairing confirmation as if the user had tapped “Allow.” The Windows exploit is a denial of service. A few packets crash and restart any machine that answers HTTP/3 through the native Windows stack, and Anthropic estimates several million such servers are listening on the internet. Claude found that bug in the binary: it reverse-engineered the executable with Ghidra and rebuilt the relevant logic as an annotated C program before hunting for flaws.
Our analysis
What we checked
On October 28 we went through Google’s 2026 Chrome stable-channel release notes, the kernel’s CVE record and fix commit for CVE-2026-98432 and Microsoft’s advisory for CVE-2026-33096. We also ran the security-guidance plugin’s pattern hook once on a synthetic file.
Anthropic’s Chrome number holds, but it was already a snapshot on the day it was published: 62 is the tally at the first Chrome 156 build, and the update a week later brought it to 80. The Windows and Linux bugs are real and fixed. Microsoft, though, files the Windows bug under HTTP.sys and rates it Important, and the kernel’s record describes the Linux bug in wider terms than Anthropic’s exploit condition.
| Anthropic’s post | What we found |
|---|---|
| 62 Chrome CVEs patched as of Chrome 156 | Exactly 62 Anthropic-credited CVEs through 156.0.8078.47 (October 20). 156.0.8078.73 added 18 more (4 Critical, 14 High). We infer that credits to Xinyang Ge without the “(Anthropic)” suffix are Anthropic’s. |
| All rated critical or high | Our 62: 14 Critical, 47 High, 1 Low (CVE-2026-106269, use-after-free in CSS). Anthropic’s own set could differ. |
| Four-bug Chrome chain | All 80 were reported between September 4 and October 21. Only 6 are in V8, and 50 are use-after-free. CVE-2026-103630, a FedCM use-after-free fixed October 1, is likely the chain’s last step (our inference). No Anthropic-credited CVE is labelled CacheStorage. |
| Linux servers with a multipath TCP listener | The kernel describes an IPv4 TCP listener race and never mentions MPTCP. CVSS 9.8, present since 4.4, fixed by Eric Dumazet on October 1. Exploitability without MPTCP is documented nowhere, and we did not test it. |
| Windows kernel QUIC bug | Listed as “HTTP.sys Denial of Service Vulnerability”: Important, CVSS 7.5, fixed April 14. Co-credited to Microsoft’s WARP & MORSE teams. |
| Plugin for PR-time scanning | The post’s link returns 404. Version 2.0.17 hooks session start, prompt submit, file edits, Bash and Stop inside Claude Code, and none of those hooks fires on a pull request. In one run it flagged yaml.load and missed a hardcoded AWS key in the same file. |
Builds, registry keys, a plugin
For Chrome, check the full version string, since the major number alone isn’t enough. The only earlier Chrome 156 build, 156.0.8078.47, lacks the last 18 fixes.
On Windows, CVE-2026-33096 affects 13 SKUs across Windows 11 23H2, 24H2, 25H2 and 26H1, Windows Server 2022, Server 2022 23H2 and Server 2025. Server 2019 and older are not on Microsoft’s list. The April update fixes the bug, and Microsoft also documents a mitigation: delete the EnableHttp3 and EnableAltSvc registry values if they are present.
Anthropic says it found more bugs by running many scanning agents in parallel with different strategies, from per-function checks to threat-model reviews, and by running variant analysis on every confirmed finding. To keep new bugs from entering a codebase, it recommends scanning during development. For Claude Code users, that means the security-guidance plugin in the anthropics/claude-plugins-official repository. Security teams can apply to the expanded Cyber Verification Program, which launched October 6 with three tiers. It offers qualifying professionals reduced blocking classifiers and access to Claude Opus 5.5, Sonnet 5.5 and Mythos 5.1.
On Linux, the CVE-2026-98432 fix is in 5.10.272, 5.15.223, 6.1.190, 6.6.159, 6.12.113, 6.18.56, 7.2.10 and mainline 7.3, which covers every maintained line kernel.org lists. For the three oldest LTS lines, those fixed versions are the current releases, out since October 15.
Up next
ClaudeClaude Codeanalysis
Claude Code 2.1.312 stops overwriting big CLAUDE.md files
A huge agent rules file in a repo subfolder could get overwritten by Claude Code after it read only the start. In 2.1.312 Claude has to read it all or edit in place. How many lines long is your biggest agent rules file?
Related stories
Comments
No comments yet. Be the first.
Join the conversation
Sign in to leave a comment. It appears after moderation.

