Anthropic's unreleased AI model hacked Android, Chrome, Windows, and Linux systems

Anthropic says an unreleased AI model independently developed five exploit chains targeting Chrome, Safari, Android, Windows, and Linux, with all reported vulnerabilities now patched and no evidence of real-world attacks. The company’s Frontier Red Team said the model linked four Chrome flaws to escape the browser’s sandbox and run code, demonstrated by opening a Mac Calculator app after a malicious page loaded. Other attacks included extracting data from a locked Android phone via USB, crashing Windows servers, and gaining root access on Linux systems under specific configurations. Anthropic has not named the model or detailed human contributions at each stage. The company frames the findings as evidence AI could help identify and fix vulnerabilities before attackers exploit them, while also highlighting risks of AI turning separate flaws into complete attacks.