Anthropic drops Claude Mythos, autonomous security model
Claude Mythos Preview (codename "Capybara") is a new high-capability frontier model from Anthropic designed to autonomously identify and exploit critical software vulnerabilities. Launched as the centerpiece of Project Glasswing, the model is restricted to a defensive coalition including AWS, CrowdStrike, and Google to proactively secure critical infrastructure before AI-powered offensive threats escalate.
The model successfully identified thousands of vulnerabilities, including a 27-year-old remote crash flaw in OpenBSD and critical bugs in the Linux kernel. Access is strictly gated through Project Glasswing, signaling a shift toward "restricted research previews" for models deemed too dangerous for general availability. The $100M in usage credits for partners indicates a massive commitment to establishing AI-augmented defense as the new industry standard. This marks the debut of the "Mythos-class" (Capybara) tier, which reportedly surpasses Claude 3.5 Sonnet and Opus in agentic coding and complex reasoning.
DISCOVERED
4d ago
2026-04-08
PUBLISHED
4d ago
2026-04-08
RELEVANCE
AUTHOR
Wes Roth