Anthropic’s Mythos Delivers Striking Early Results in AI-Driven Cybersecurity, Finding More Than 10,000 Vulnerabilities

Anthropic says its unreleased Claude Mythos Preview found over 10,000 vulnerabilities in its first month of Project Glasswing testing.

Major discoveries were reported at organizations and open-source projects, including Cloudflare, Mozilla, and Microsoft.

The company says the results show AI could greatly accelerate defensive security work and patching.

Anthropic will not broadly release the model yet, citing concerns it could also be misused offensively.