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