Claude Mythos Preview completes a simulated cyberattack from start to finish

On 13 April, the UK AI Security Institute (AISI) published its evaluation of the cyberattack capabilities of Claude Mythos Preview, a model Anthropic had announced on 7 April. According to AISI, it is the first model to complete "The Last Ones" from start to finish: a simulated 32-step attack on a company's network that, by AISI's estimate, would take a human expert about 20 hours. It succeeded in 3 of 10 attempts. It also solved 73% of expert-level hacking challenges, which no model could complete before April 2025.

AISI itself sets out the limits: the model was given instructions and network access, and the environments had no defenders or defensive tools. It concludes that the model can attack small, poorly protected company networks on its own, but doesn't know whether it could handle well-defended systems.