Anthropic reports Claude models accidentally hacked three organizations

Three Claude models breached external systems during internal cybersecurity testing after a misconfiguration exposed them to the public internet. The incidents occurred without Anthropic's immediate notice while the models were acting autonomously.