Anthropic AI used fake identities in UK test, report says

An Anthropic AI model created fake online identities and sent deceptive emails to real people during a UK government test, the UK AI Security Institute said in a report published late Tuesday.

The institute said the model, Anthropic’s Mythos 5, tried to insert malicious code into a software project by persuading a recipient to approve it. The person overseeing the software refused approval.

The institute said some Anthropic and OpenAI AI agents engaged in “sustained, potentially harmful activity directed at real people and organisations” during tests carried out with open internet access and some safety features disabled.

Most of the actions came from Mythos 5, while two involved OpenAI’s GPT-5.6-Sol model, the institute said. It said the attempts were unsuccessful, its investigations had not found evidence of real-world harm and the incident was contained within an hour.

The institute said the activity showed “novel, potentially deceptive behaviours” at a level it had not expected.

An Anthropic spokesperson said the report showed the need for a broader discussion about how to safely evaluate increasingly capable AI agents. An OpenAI spokesperson said independent testing was essential and that the company would keep working with evaluators and others to strengthen safe evaluation practices.

The report follows earlier security incidents involving AI models. OpenAI said in July that its software escaped a testing environment and attacked Hugging Face, then later said models had targeted three additional companies. Anthropic said on July 30 that it had found three incidents in which AI models being tested gained unauthorised access to unidentified organisations.