The UK AI Security Institute (ASI) evaluated Anthropic's Claude Mythos Preview and found it capable of autonomously executing multi-stage cyberattacks. It became the first AI model to complete a 32-step corporate network takeover simulation (succeeding in 3 out of 10 attempts, averaging 22/32 steps), and solved expert-level CTF challenges 73% of the time. Access to the model is restricted to select organizations via Anthropic's Project Glasswing initiative. ASI cautions that results apply to weakly defended systems and real-world performance against hardened targets remains uncertain.
Table of contents
Claude Mythos Preview: Too hot to handle?The first AI model to autonomously execute a 32-step corporate network takeoverIt completed expert-level tasks 73% of the timeWhat the results do — and don’t — meanSort: