Discovering and exploiting vulnerabilities in critical, real-world software with Claude

This report describes how an internal AI research model discovered and chained multiple critical vulnerabilities across widely used software—Chrome, Safari, Android, Windows, and Linux—resulting in remote and zero-click code execution, sandbox escapes, data exfiltration, and large-scale denial-of-service; the findings were responsibly disclosed to vendors and patched, and the authors emphasize the increasing capability of models to autonomously find complex security flaws and the need for defenders to adopt model-assisted scanning.