Anthropic's Claude finds 62 bugs in Chrome and builds working exploit chains
An unreleased internal research version of Anthropic's Claude model independently chained multiple vulnerabilities across major browsers and operating systems.
Anthropic detailed how its internal research model discovered and chained critical vulnerabilities across major browsers and operating systems, including Google Chrome, Apple's Safari, Android, Windows, and Linux. The unreleased model autonomously found bugs and developed working exploit chains after an initial task setup.
In Google Chrome, the model linked memory corruption in V8, an ASLR bypass, a CacheStorage race condition, and a FedCM use-after-free error. This four-step chain broke out of the browser sandbox to execute the Calculator app on macOS with default settings. Anthropic reported 62 CVE-indexed vulnerabilities to Google, all fixed by version 156.
For Safari, Claude bypassed the Same-Origin Policy without executing arbitrary code by exploiting use-after-free and IPC flaws to modify memory and read cookies, history, and active banking or email sessions. Apple has since issued patches. On Android, the model analyzed open-source code and Google Play services binaries to extract photos from a locked Google Pixel 9 over USB without a passcode.
The model also targeted operating system kernels. Using reverse engineering tools like Ghidra on a Windows binary, it found a QUIC protocol heap-reading bug (CVE-2026-33096) causing remote denial-of-service on HTTP/3 servers. In Linux, it discovered an MPTCP use-after-free race condition (CVE-2026-98432) that bypassed KASLR to execute commands with root privileges.
Source: habr
Comments
Help shape the next version: Add context or suggest a correction. AI review can add points toward a rewrite. Reviews and updates may take time; a full meter does not guarantee a new version.