As per a report by MediaNama, Anthropic's Frontier Red Team reported that an unreleased internal research model produced complete exploit chains against five targets. In its research post, dated October 27, Anthropic describes a four-stage chain on Chrome for macOS that ran code outside the browser sandbox, with 62 CVEs assigned and patched as of Chrome 156, all rated critical or high; a one-click Safari attack that bypassed the same-origin policy to reach browsing history, cookies and authenticated requests, which Anthropic says Apple acknowledged the day it was reported; and a lock-screen bypass on a Google Pixel 9 that extracted photos over USB, which Google has addressed. It also describes a remote crash in the Windows kernel's QUIC implementation (CVE-2026-33096) and a zero-click remote root exploit through a race condition in Linux multipath TCP (CVE-2026-98432), patched upstream and backported to stable kernels. Anthropic says the Linux exploit took 48 minutes to leak a kernel pointer and five more to reach command execution.

Anthropic says it reported everything under coordinated disclosure, giving vendors 90 days to patch and a further 45 days before it discusses details, and that the post covers only vulnerabilities that have since been resolved. It gives technical details only for the Linux bug. MediaNama notes that Anthropic does not disclose the number of attempts, the computing resources used, the false positives or the targets on which the model failed, or what the prompts contained and how much human guidance the model received, and that the findings have not been independently verified. Anthropic also says it has expanded its Cyber Verification Program for qualifying security professionals.