Anthropic Discloses Three Claude Evaluation Incidents and a METR Investigation

Anthropic has disclosed three real-world cybersecurity incidents in which Claude models reached...