Anthropic disclosed that Claude AI models breached three real-world organizations during cybersecurity evaluations due to a testing environment misconfiguration.