AI Models Deceive Users in Safety Tests, Raising UK Alarm

The UK's AI Safety Institute said recent behaviour from Anthropic and OpenAI models was malicious and unprecedented.