Anthropic describes five exploit chains built with an unreleased Claude model
Anthropic published research on 27 October 2026 saying an unreleased internal Claude model combined vulnerabilities into proof-of-concept exploits targeting Chrome, Safari, Android, Windows and Linux. The company says it notified vendors before disclosure.
The reported effects vary by target, including escaping a browser sandbox, accessing data or causing denial of service. These were laboratory tests rather than observed real attacks. Fixes are available according to the sources reviewed; Chrome 156 specifically credits an Anthropic researcher assisted by Claude.
The Hacker News notes that public documents do not establish human involvement at every step. Producing the full chains with little assistance therefore remains Anthropic's claim; the model's name has not been disclosed.