Cloudflare tested Anthropic's security-focused AI model Mythos Preview across more than 50 of its own code repositories as part of Project Glasswing. The model can chain multiple small vulnerabilities into working exploits and prove they're real by writing, compiling, and running proof-of-concept code on its own.
Earlier frontier models found similar individual bugs and sometimes delivered solid analysis, according to Cloudflare CSO Grant Bourzikas. But they fell short at stitching the pieces together, leaving chains unfinished and the question of actual exploitability open. Mythos Preview produced fewer speculative findings, clearer steps to reproduce issues, and needed less human follow-up to reach a fix-or-dismiss decision.
Cloudflare stresses that a single agent isn't enough. The company built a multi-stage harness with up to 50 parallel agents and adversarial review, where a second agent tries to disprove each finding. The company also warns: these same capabilities will be available to attackers, too.
AI News Without the Hype – Curated by Humans
Subscribe now
