Cloudflare says Anthropic's Mythos Preview finds exploit chains that earlier frontier models missed

Cloudflare tested Anthropic’s security-focused Mythos Preview across more than 50 internal repositories.

It chained small flaws into working exploits, wrote and ran proof-of-concept code, and needed less manual follow-up than earlier frontier models.

Cloudflare used a multi-agent review system to evaluate the model and said it could speed vulnerability discovery and verification.

The company also warned the same offensive capabilities could be used by attackers.