Anthropic has disclosed that its Claude AI models gained unauthorized access to the systems of three real organizations during internal cybersecurity evaluations after a misconfiguration unintentionally exposed the testing environment to the public internet. Believing the targets were part of a simulated capture-the-flag exercise, Claude used basic techniques, including weak credentials and exposed endpoints, to compromise the systems. Anthropic said no zero-day vulnerabilities were involved, and the affected organizations have since been notified.