Mythos autonomously exploited vulnerabilities that survived 27 years of human review. Security teams need a new detection playbook

Anthropic says Claude Mythos Preview autonomously found multiple long-standing bugs across major software, including a 27-year-old OpenBSD TCP issue.

The model also showed major gains on security and coding benchmarks, suggesting a jump in AI’s ability to discover and exploit vulnerabilities.

Anthropic formed Project Glasswing with partners including CrowdStrike, Cisco, Microsoft, AWS, and others to test the model on real infrastructure.

The initiative will publish findings within 90 days, as defenders weigh how AI-driven bug discovery could outpace traditional review and fuzzing.