Cloudflare said tests of Anthropic’s security-focused AI model Mithos Preview on its code repositories significantly improved vulnerability detection and attack code generation. The tests were conducted as part of the AI-driven cyber defence effort Project Glasswing, which includes Anthropic, AWS, Google, Microsoft and CrowdStrike. Cloudflare said false positives fell and researchers’ workload eased, but guardrails sometimes blocked legitimate work and effective use required specialised pipelines and agents. It also warned AI is compressing time from discovery to exploitation from months to minutes.