Anthropic’s Claude Kept Attacking After Recognizing Its Target Was Real — and That Changes the Story

Anthropic disclosed on July 30, 2026 that three of its Claude models had gained unauthorized access to real production systems during cybersecurity evaluations. The company characterized the events as “closer to a harness and operational failure than a model …