Frontier AI Can't Hack Corporate Networks? Claude Mythos Just Did It in 20 Hours.
A 32-step corporate network attack. 20 hours of human red-team work. Completed start-to-finish by an AI. Three times out of ten.
The UK AI Security Institute (AISI) published its independent evaluation of Claude Mythos Preview today. The results are the first independent confirmation of what people inside Anthropic have been quietly terrified about since February.
Key Takeaways
- ✓AISI published first independent evaluation of Claude Mythos — model completed 32-step TLO attack autonomously 3 of 10 attempts, average 22 steps
- ✓73% success on expert-difficulty capture-the-flag challenges that scored 0% twelve months ago — the trajectory line is the story
- ✓Project Glasswing access tier prices Mythos at $25/$125 per million tokens (5x Opus 4.7) to create economic friction
- ✓Test ranges lacked EDR, active defenders, and incident response — Mythos can attack weakly-defended networks, not hardened enterprise environments yet
- ✓OpenAI confirmed same-day that it has a restricted cyber model ready to release through a similar consortium structure — arms race is no longer theoretical
Skila AI Editorial Team
The Skila AI editorial team researches and writes original content covering AI tools, model releases, open-source developments, and industry analysis. Our goal is to cut through the noise and give developers, product teams, and AI enthusiasts accurate, timely, and actionable information about the fast-moving AI ecosystem.
About Skila AI →Related Resources
Weekly AI Digest
Get the top AI news, tool reviews, and developer insights delivered every week. No spam, unsubscribe anytime.
Join 1,000+ AI enthusiasts. Free forever.