Booz Allen Hamilton's first-ever Cyber Weapon Index found that Anthropic's Claude Mythos was the only AI model, out of 18 tested, to autonomously complete an entire cyber attack "kill chain" from vulnerability discovery through to full network compromise. The consultancy warns that most of the other 17 US and Chinese models are likely to reach a similar level of weaponisation within six months, and describes mainstream AI-enabled attacks by ransomware gangs and state-backed hackers as "imminent." It is urging the US government to impose sector-specific resilience deadlines on critical infrastructure and to build up both offensive and defensive AI capabilities.

The index scored 18 models—nine American, nine Chinese—on their ability to identify vulnerabilities and execute intrusions, both with and without stolen credentials. Claude Mythos topped the table with a score of 80, gaining administrator-level control in every attempt when given credentials and still achieving full domain compromise without them, reportedly working out escalation routes independently rather than following a scripted plan. It was followed by xAI's Grok-4.5 (49) and OpenAI's GPT-5.6 Sol (46); three models (Grok-4.5, Meta's Muse Spark 1.1, and Z.ai's GLM-5.2) also reached full domain control, while four others achieved lateral movement. Separately, OpenAI said its unreleased Astra model has crossed a "critical" cybersecurity risk threshold for exploiting zero-day flaws.