Cloudflare validates Anthropic’s unreleased AI model “Mythos”: what its capabilities revealed in a vulnerability investigation

Cloudflare tested Anthropic’s unreleased security-focused AI model, Claude Mythos Preview, on more than 50 of its own and related open-source repositories.

The model helped uncover chained attack paths across multiple flaws and even generate proof-of-concept code.

The results suggest advanced AI could speed up defensive vulnerability research, but also make attack automation easier.

Cloudflare says practical use will require strong safety controls and validation infrastructure.