A cheap piece of software erased Booz Allen’s own AI threat ranking

Booz Allen tested 18 top AI models by having each autonomously breach a live corporate network, tracking code vulnerability discovery and escalation to domain-admin control. Anthropic’s Claude Mythos scored 80, gaining admin rights both with and without stolen credentials; all other models failed the credential-less scenario. Adding a dedicated attack harness markedly reduced performance gaps, showing that system-level integration, not the model alone, drives cyber-risk.