By Hamza Faisal
Anthropic has revealed that an unreleased internal Claude research model found critical flaws in some of the world's most widely used software — and then chained them together into five working exploits. The company's Frontier Red Team published the findings on October 27, describing attacks against Google Chrome, Apple Safari, Android, Windows and the Linux kernel, each built with what Anthropic calls minimal human intervention after an initial prompt.
None of the attacks were seen in the wild. Anthropic says every flaw was reported to the affected vendor under coordinated disclosure and has since been patched, so the research is a warning about what AI models can now do rather than a live threat. But the demonstrations mark a shift: finding a single bug is one thing, but turning several bugs into a finished exploit chain has traditionally required specialist human teams.
A one-click escape from Chrome's sandbox
The most detailed chain targeted Chrome. Starting from a single page load on a default macOS installation, the model linked four vulnerabilities: a memory-corruption flaw in V8, Chrome's JavaScript engine; an escape from V8's memory sandbox that defeated address randomisation and took over the renderer process; a race condition in CacheStorage that mapped the privileged browser process's memory; and a use-after-free bug in FedCM, Chrome's sign-in feature, that completed the sandbox escape. The proof of concept opened the Calculator app within seconds of the page loading, with no further clicks.
Anthropic says Google assigned 62 CVE identifiers to findings from this research, all rated critical or high, including 13 sandbox escapes, and patched them as of Chrome 156. Chrome's release notes credit an Anthropic researcher, assisted by Claude, for several of the fixes, though they do not account for all 62 — a gap The Hacker News has noted.
Safari, Android and Windows chains
The Safari chain never ran code at all. Using a use-after-free flaw in JavaScriptCore, the model worked around WebKit's Gigacage memory-isolation defence and exploited the sandbox's messaging mechanism to read on-disk browsing history, cookies and session data, then flipped a single in-memory flag to disable the same-origin policy — the rule that stops one website reading another site's data. Apple acknowledged the report and shipped a fix.
On Android, the model combined source review with reverse-engineered Play Services binaries to pull photos and files from a locked Pixel 9 over USB, bypassing the pairing authorisation and six-digit verification checks. On Windows, it found an out-of-bounds read in the HTTP/3 network stack that could repeatedly crash a server remotely, though it could not run code or read stored data; Microsoft fixed it as CVE-2026-33096.
A zero-click root shell on Linux
The Linux chain was the most severe. Exploiting a race condition in the multipath TCP network path, the model took 48 minutes to leak a kernel pointer and five more minutes to execute commands as root on a server running a multipath TCP listener — with no login and no user interaction. The flaw was patched upstream as CVE-2026-98432, though it required that specific networking configuration and would not work against every Linux system.
Anthropic has not named the model, said when it might be released, or fully explained how much researchers contributed at each step — caveats that matter when judging how autonomous the work really was. Its conclusion for defenders is blunt: exploit chaining used to need specialist teams, and models can now do much of it, so organisations should scan their own code earlier and more often, and users should keep browsers, phones and kernels updated.



