The UK AI Security Institute published independent evaluation of Mythos

The UK AI Safety Institute published an independent evaluation of Claude Mythos Preview, finding the model represents a major step forward in autonomous cyberattack capability. The evaluation marks the first public confirmation that Anthropic is developing a model under the Mythos name, which does not match existing Claude 3, 3.5, 3.7, or Claude 4 lines and may be a specialized variant for extended reasoning or agentic use cases. The UK AI Safety Institute, which has evaluated frontier models since late 2023, uses uplift testing to measure whether AI bridges the gap between malicious intent and actual execution skills under realistic conditions. While previous evaluations found current models provide some uplift to less-skilled attackers on basic tasks without dramatically accelerating skilled threat actors, the Mythos evaluation suggests meaningful advancement in autonomous offensive cyber capabilities. The UK is working toward making such pre-deployment evaluations mandatory under upcoming AI regulation.

UK AI Safety Institute's evaluation of Claude Mythos Preview

• ▪Previous UK AI Safety Institute evaluations found that current models provide some uplift to less-skilled attackers on basic tasks
• ▪The UK AI Safety Institute evaluated Claude Mythos Preview as a pre-release version of what appears to be a new Claude model family
• ▪The UK AI Safety Institute is the UK government body responsible for testing frontier AI models before public release
• ▪The UK AI Safety Institute works directly with Anthropic, OpenAI, Google DeepMind, and others on model evaluations
• ▪UK AI Safety Institute evaluations found that AI models are useful for reconnaissance and planning phases of cyber attacks
• ▪Pre-deployment access to frontier AI models was initially agreed to voluntarily by companies working with the UK AI Safety Institute
• ▪The UK AI Safety Institute published an evaluation of Claude Mythos Preview's cyber capabilities
• ▪Previous UK AI Safety Institute evaluations found that current models do not dramatically accelerate the capabilities of already-skilled threat actors
• ▪The UK AI Safety Institute has been running evaluations on frontier models since late 2023
• ▪The UK AI Safety Institute evaluation of Claude Mythos Preview focused specifically on whether the model provides meaningful uplift to people attempting offensive cyber operations

First public confirmation of Anthropic's Mythos model line

• ▪Claude Mythos does not match any existing Claude line including Claude 3, 3.5, 3.7, or Claude 4

Perspective of UK AI Safety Institute

• ▪The UK AI Safety Institute evaluation found that Claude Mythos Preview marked a major step forward in autonomous cyberattack capability

Perspective of Anthropic

• ▪Anthropic's development of Claude Mythos Preview represents exploration of new model architectures beyond the existing Claude 3 and 4 families