IC Technology & AI — 2026-04-15
UK AI Safety Institute Finds Claude Mythos Preview Capable of Autonomous Network Takeover Attacks
The UK AI Safety Institute released evaluation results showing Anthropic's Claude Mythos Preview can autonomously execute multi-stage network takeover attacks on vulnerable systems, succeeding 73% of the time on expert-level cybersecurity tasks that no AI model could complete before April 2025. The findings prompted Anthropic to launch Project Glasswing, deploying the model to identify thousands of zero-day vulnerabilities across major operating systems and browsers for defensive purposes. U.S. and UK cyber agencies are grappling with the dual-use implications of AI systems that can independently discover and exploit software vulnerabilities at scale.
Analysis
The 73% autonomous success rate, on tasks no prior model could complete, establishes an offensive cyber capability baseline that intelligence agencies on both sides are almost certainly already working to operationalize, with export control frameworks carrying no existing mechanism to restrict AI at the vulnerability-discovery layer. Converging reporting confirms Anthropic redirected the same capability toward defensive zero-day discovery under Project Glasswing, but that reframing offers no restriction: any actor with model access can weaponize it offensively. The benchmark may overstate real-world threat if evaluation systems were less hardened than production networks, where the gap between laboratory exploitation and live operations remains material even for human operators; the interagency implication holds either way. U.S.-UK agencies will very likely issue new classification guidance for AI model evaluation results within the coming months; that guidance will lag rather than contain a capability that is already public.
The 73% autonomous success rate, on tasks no prior model could complete, establishes an offensive cyber capability baseline that intelligence agencies on both sides are almost certainly already working to operationalize, with export control frameworks carrying no existing mechanism to restrict AI at the vulnerability-discovery layer. Converging reporting confirms Anthropic redirected the same capability toward defensive zero-day discovery under Project Glasswing, but that reframing offers no restriction: any actor with model access can weaponize it offensively. The benchmark may overstate real-world threat if evaluation systems were less hardened than production networks, where the gap between laboratory exploitation and live operations remains material even for human operators; the interagency implication holds either way. U.S.-UK agencies will very likely issue new classification guidance for AI model evaluation results within the coming months; that guidance will lag rather than contain a capability that is already public.
3 sources
- Our evaluation of Claude Mythos Preview's cyber capabilities -
UK AI Safety Institute - Here's how cyber heavyweights in the US and UK are dealing with Claude Mythos -
CyberScoop - Testing reveals Claude Mythos's offensive capabilities and limits -
Help Net Security