Claude Mythos alone completes AI cyberattack chain

Booz Allen Hamilton tested 18 frontier AI models as autonomous cyber attackers and found only one, Anthropic's Claude Mythos, could complete a full attack chain without human help. In live network tests, the model broke in, escalated to administrator-level control, and achieved full domain compromise on its own, scoring 80 on the firm's new Cyber Weapon Index, well ahead of Grok-4.5 (49) and GPT-5.6 Sol (46) for second place.

The gap between Claude Mythos and the rest of the field is not as wide as the top score suggests. Three other models reached domain access, four achieved lateral movement inside the network, and all but one managed initial access. Booz Allen's own read is that the remaining models are likely to close the gap to Claude Mythos' current capability within six months, not that this stays a one-model problem.

For teams building on AI coding agents or exposing agent-accessible infrastructure, the test scenarios matter more than the leaderboard: models are being scored on independently identifying escalation paths rather than following a scripted attack sequence, the same autonomous, adaptive behavior that makes agentic coding tools useful for legitimate work. Credential hygiene, network segmentation, and treating any agent-reachable system as a potential attack surface are no longer defenses against a hypothetical, they are defenses against a benchmarked capability that already exists in a model available today.