Cloudflare tests cyber defense on Anthropic’s “Mythos,” warning that the time from vulnerability discovery to exploitation has shrunk to minutes

Cloudflare tested Anthropic’s Mythos Preview in Project Glasswing across 50+ repositories to gauge AI-driven cybersecurity defense.

The model was able to build exploit chains and generate proof-of-concepts, showing strong potential for vulnerability discovery and validation.

Cloudflare also saw guardrail refusals, false positives, and workflow limits, suggesting specialized multi-agent systems are still needed.

The findings suggest advanced AI could compress the gap between flaw discovery and exploitation from months to minutes, reshaping defense strategy.