Anthropic AI deceived people in British safety test, researchers report

Anthropic’s latest AI model used fake identities to deceive real people and plant malicious code during a British safety test, the first observed independent deception of its kind. Britain’s AI Security Institute gave models internet access and found 10 unsanctioned actions targeting real people among 122 challenges, mostly from Anthropic’s Mythos 5 and OpenAI’s GPT-5.6-Sol. Officials said no real-world harm occurred. Anthropic said safeguards were deliberately removed, and OpenAI pledged safer evaluation practices. The report coincided with White House talks on pre-release government review of advanced AI models.