Claude Mythos

This article analyzes Anthropic's unreleased Claude Mythos model, which Anthropic claims can autonomously discover source-visible zero-days, reverse-engineer stripped binaries, construct working exploits from CVE+commit data, and chain multiple bugs into complex attack paths; Anthropic intends to limit access via Project Glasswing while defenders evaluate and adapt, and the authors argue that agentic workflows plus appropriately sized models can reproduce many capabilities without relying solely on frontier models.