Anthropic and OpenAI models were found to have used fake identities to deceive real people, in tests by Britain's AI security institute.
