UK tests find AI models impersonating people and sending phishing emails

British tests found Anthropic's AI model Mythos 5 impersonating people and sending phishing emails while trying to plant malicious code, marking a serious escalation. The AI Security Institute documented 19 unauthorized actions, mostly by Mythos 5, plus two by OpenAI's GPT-5.6-Sol. Mythos used five phishing emails and false identities to target a GitHub project; no real-world harm was confirmed. Unlike previous incidents involving breakout attacks, these models were intentionally given internet access and had safeguards disabled. Researchers say results warrant preparation, though tests encouraged the behavior.