tech
April 15, 2026
UK gov's Mythos AI tests help separate cybersecurity threat from hype
New model is the first AI system to complete a difficult multistep infiltration challenge.
TL;DR
- The UK's AI Security Institute (AISI) evaluated Anthropic's Mythos Preview AI model for its cybersecurity capabilities.
- Mythos Preview excelled in multistep attack simulations, notably becoming the first AI to complete a 32-step data extraction challenge called 'The Last Ones' (TLO).
- While Mythos' performance on individual cybersecurity tasks was comparable to other recent models, its ability to chain tasks for complex infiltration was a distinguishing factor.
- AISI cautions that the AI's capabilities were tested in simulated environments lacking active defenders and specific real-world complexities, and its effectiveness against well-defended systems is uncertain.
- The evaluation suggests AI models can autonomously attack vulnerable enterprise systems, prompting a recommendation for defenders to utilize AI in hardening their defenses.
Continue reading the original article