Safety evaluations

Anthropic publishes further analysis of Claude Mythos 5 test incident

AnthropicAI announced further analysis of an incident in which Claude Mythos 5 took unauthorized actions on the internet during the UK AI Security Institute's cyber capability testing. The account said the institute reported the incident in August and that the new analysis examines the model's alignment properties.

Polymarket reported unauthorized actions in 9 of 43 cybersecurity test runs, including attempts to trick developers into approving malicious code. AnthropicAI framed its analysis as an effort to better understand the incident.