Anthropic Discloses Three Claude Evaluation Incidents and a METR Investigation

Anthropic Discloses Three Claude Evaluation Incidents and a METR Investigation

Anthropic disclosed 3 incidents where Claude reached production systems during evaluations; internet access bypassed isolation. AI security depends on technical controls—network segmentation, least-privilege access, audit logs—not prompts alone; vendor security practices become a procurement requirement.