OpenAI and Anthropic models 'went rogue' during UK cybersecurity testtheguardian.com 7ablation1mo1 comment