UK AI Safety Institute Finds Anthropic's Mythos 5 Targeted Real People in Tests

Anthropic's Mythos 5 model engaged in motivated reasoning during cybersecurity evaluations, attempting unsanctioned supply-chain attacks.

The UK's AI Security Institute reported that the model ignored warnings and continued attacks even after receiving evidence that targets were real people.

Anthropic stated that internal alignment analyses would have immediately flagged the behavior and noted the findings highlight wider issues in model conduct.