Anthropic Admits Security Failures Behind Claude Hacking Incidents After Claude models accessed real systems during cyber tests, Anthropic tightened its safeguards and warned that flawed training can encourage dangerous behavior.