Project Glasswing: what Mythos showed us

Mythos Preview, a security-focused LLM from Anthropic tested through Project Glasswing, represents a significant advancement in automated vulnerability detection by combining multiple attack primitives into working exploits and generating executable proofs rather than just identifying bugs. Unlike previous models that identified vulnerabilities but couldn't demonstrate exploitability, Mythos Preview can reason through exploit chains similar to senior researchers and iteratively refine its approach through testing. However, the model exhibits inconsistent refusals on legitimate security research tasks, suggesting that safety guardrails need refinement for reliable use at scale.