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

AI Security Institute says tools engaged in potentially harmful activity and incident reveals new type of risk