Anthropic says its AI linked software flaws into five working attacks

Anthropic says one of its unreleased Claude models did more than spot software flaws. It linked them together into five working attacks on Chrome, Safari, Android, Windows and Linux. The company published the research on Oct. 27. All of it happened in controlled tests, not in real-world attacks.

What changed

Security people call a set of linked flaws an exploit chain. Each flaw alone may be blocked by other protections, but together they can get past several layers. Anthropic says Claude studied code, reverse-engineered software and tested attacks after being prompted.

In Chrome, the model linked four flaws. A malicious webpage then launched the Calculator app on a Mac without any further action. In the Linux demo, the attack ran commands with root-level access, the highest level of control. Some chains began with a webpage. Others needed physical access or a particular network service.

Why it matters

Anthropic's wider Chrome research found 62 flaws, including 13 that let attackers escape the browser's protective sandbox. Fixes are included by Chrome 156. The four-flaw chain was one demonstration within that larger effort.

The worry is speed. If AI helps attackers build these chains faster, the gap between a flaw being found and being used could shrink. Installing updates promptly on your phone, computer and browser is the practical response.

The report excerpt does not say whether the model will be released or how fixes are rolling out on every platform.