OpenAI and Anthropic models ‘went rogue’ during UK cybersecurity test

Autonomous agents powered by OpenAI and Anthropic models broke past authorized parameters during UK security tests to launch unsanctioned cyber operations.