Anthropic: unreleased Claude model found critical bugs in Chrome, Safari, Android, Windows and Linux

On 27 October 2026 Anthropic’s Frontier Red Team published a report on using an unreleased internal research model to find critical vulnerabilities in widely used browsers and operating systems and turn them into working attack chains “with little to no user interaction”. The company says every issue was reported to the vendor under coordinated vulnerability disclosure (90 days to patch, plus 45 days before further details are discussed).

Five disclosed chains (high level)

• Chrome: a multi-bug chain that could run code outside the browser sandbox after a user visited a malicious page. Anthropic says Google has assigned and patched 62 CVEs as of Chrome 156 (critical or high), including 13 sandbox escapes.
• Safari: a chain that could bypass the same-origin policy and access data from other sites where the user is signed in. Anthropic says Apple has shipped a fix.
• Android: chained flaws that, on a locked Pixel 9 in Anthropic’s test, allowed personal data to be pulled over USB. Google has addressed the issue, Anthropic says.
• Windows: a remotely triggerable flaw in the native HTTP/3 (QUIC) stack that could crash listening servers; Microsoft assigned CVE-2026-33096.
• Linux: a race condition in the multipath TCP networking path that Anthropic says enabled remote code execution as root without credentials; patched upstream as CVE-2026-98432 and backported to maintained kernels.

Context and limits

Anthropic frames the work as evidence that frontier models can now discover and combine multiple bugs into a verified exploit with limited human prompting, and urges defenders to scan their own code (including at pull-request time). The public post describes outcomes and vendor status; it does not publish step-by-step attack recipes for outsiders. Details below the headline figures come from Anthropic’s own announcement and have not been independently re-tested here.