Anthropic's internal Claude research model finds flaws across major systems

According to an IT之家 report on October 29, Anthropic published a research report on October 27 describing how an internal, unreleased Claude research model discovered vulnerabilities in Chrome, Safari, Android, Windows and Linux, chaining them into five fully functional attack chains. The findings were reported to the vendors through coordinated disclosure and have been fixed.

The Chrome work took the spotlight: Anthropic's report to Google produced 62 critical or high-severity CVE entries, including 13 sandbox escapes and 21 flaws that allow control of the renderer process, with the fixes shipped in Chrome 156. According to the research team, Claude located the vulnerabilities and wrote exploit code on its own after being given an initial prompt; the Chrome proof-of-concept opens a calculator application in the default macOS environment within seconds of a page loading. On Safari, the chain Claude uncovered can bypass the same-origin policy, letting a malicious site read data from accounts the user is signed into; Apple confirmed and shipped the fix the same day. The Android attack requires physical access to the device, and can extract photos and other personal data from a Pixel 9 while it is locked, via USB; Google said it has addressed the issue.

The report states that this internal Claude research model has not been publicly released. Coverage so far notes only that Windows and Linux were also included in the attack chains with the vulnerabilities reported through the disclosure process; details on those two platforms have not been disclosed.