Project Glasswing: what Mythos showed us

Cloudflare evaluated Anthropic's Mythos Preview LLM against dozens of internal repositories and found it significantly improves vulnerability discovery by chaining primitives into working exploits and generating proofs-of-concept; the report describes the model's inconsistent refusals, signal-to-noise challenges, and a multi-stage harness (recon, hunt, validate, trace, etc.) used to increase coverage and reduce false positives, while emphasizing the need for additional safeguards and controlled use.