Anthropic Model Builds Exploits for Chrome, Safari, Windows and Linux

An unreleased model from Anthropic found security flaws in Chrome, Safari, Android, Windows and Linux, then combined them into five working exploit chains, the company said in a research post. The tests occurred in a lab, no real attacks were observed, and the affected vendors patched the reported flaws.

The Chrome exploit combined four flaws to escape the browser sandbox and run code on a Mac after a page loaded. Anthropic says it reported 62 Chrome flaws that received CVE identifiers, including 13 sandbox escapes, with fixes included in Chrome 156.

Other exploit chains could bypass Safari's same origin policy, extract data from a locked Android phone over USB, crash a Windows server through HTTP/3 requests, or run commands as root on a Linux server using multipath TCP. Apple and Google credited Anthropic researchers and Claude for some fixes. Anthropic says the work needed only initial prompting, but as The Hacker News noted, it did not name the model or break down the human involvement in each chain.