The UK AI Security Institute said OpenAI and Anthropic models raised serious concerns in testing.