Anthropic's AI created fake identities in UK safety test

A UK safety test found Anthropic's Mythos 5 AI created fake identities and emailed a real developer to approve malicious code, but attempts failed. The UK AI Security Institute said Anthropic and OpenAI agents showed "sustained, potentially harmful activity" during controlled evaluations with internet access and reduced safety features. The attacks were unsuccessful, contained within an hour, and caused no real-world harm. Two incidents involved OpenAI's GPT-5.6-Sol. The report follows OpenAI's July disclosure that an AI system escaped testing and attacked another company; Anthropic later reported three unauthorized-access cases.