There is a national-security imperative to safeguard the most advanced AI models, build defensive AI that can detect and respond at machine speed, and prepare for fully autonomous cyber agents.

Frontier API models generally score highly on vulnerability discovery in synthetic tests when vulnerabilities are real, but several models fail to exploit real bugs, with Claude Mythos uniquely exploiting one unseen vulnerability.

Attack harnesses—connecting models to hacking tools and orchestration logic—dramatically boost a model’s offensive capabilities, sometimes rivaling or exceeding the best model without harnesses.

Claude Mythos from Anthropic autonomously completed the full cyber kill chain, even when using stolen credentials, and did so without human intervention.

Booz Allen conducted the first Cyber Weapon Index, evaluating 18 AI models from US and Chinese developers under identical conditions to gauge autonomous vulnerability discovery, capability development, and attack execution.

OpenAI’s Astra, not part of the tested set, reportedly reached a critical cybersecurity capability threshold, signaling strong vulnerability-discovery potential.

The report cautions that real-world impact depends on system vulnerabilities and defense-in-depth, and warns that autonomous agents could operate beyond intended mission parameters.

Several models achieved advanced stages such as full domain access, lateral movement, and credential acquisition; three models in particular (Grok-4.5, Muse Spark 1.1, GLM-5.2) reached full domain access and control.

Other models reached various advanced milestones (domain access, lateral movement, credentials) but did not complete the entire kill chain; credentials significantly expanded some capabilities.

The index merges vulnerability research score and kill chain attainment score, with Claude Mythos leading the field (CWI 80), followed by Grok-4.5, GPT-5.6 Sol, Muse Spark 1.1, and others.

Overall, the findings highlight strong offensive potential for autonomous AI in cyber operations and underscore the urgency for defensive measures and policy actions to mitigate risks.

The attack harness can markedly amplify capabilities, raising concerns that fully capable model-harness systems already exist and could challenge national control over such capabilities.