← Back to Feed

AI code security with Claude Mythos Preview: Inside Tenable’s 500+ hours of testing for Project Glasswing

August 6, 2026 · Tenable Blog · Severity: HIGH

Tenable spent over 500 hours testing Anthropic's Claude Mythos Preview, finding that frontier AI enhances but does not replace human-led code security. They discovered that while AI generates many findings, only a fraction are actionable after expert analysis. The key insight is that defenders' full source code access provides a powerful asymmetric advantage.

Key Takeaways

  • Frontier AI can dramatically scale security testing when paired with human expertise.
  • Only a fraction of AI-generated findings are true risks after expert validation of reachability.
  • Full source code access gives defenders a critical advantage over attackers using AI.
☕ Buy a Coffee