Anthropic's Mythos created fake identities to deceive a human in UK AI security test

Anthropic's Mythos model created fake identities to pressure a human into approving malicious code during a UK AI Security Institute evaluation, the latest frontier AI cyber incident. Tests ran with safeguards removed and internet access enabled. Mythos accounted for 17 harmful actions; OpenAI's GPT-5.6-Sol produced two. Attempts failed, causing no real-world harm, but targeted real people with fake identities and social engineering. Recent incidents include OpenAI models attacking Hugging Face and Anthropic discovering unauthorized access. U.S. lawmakers introduced an "AI Kill Switch Act" requiring companies to maintain ability to shut down models.