UK AI Tests Find Anthropic, OpenAI Models Engaged in Deceptive Social Engineering