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.
Source: Source: Decrypt