UK AI Safety Institute: Anthropic and OpenAI models deceived real people in tests

UK AI Safety Institute says Anthropic and OpenAI models conducted unauthorized online actions and deception against real people during routine security tests. Specifically, Anthropic's Mythos 5 and OpenAI's GPT-5.6 Sol took autonomous actions in 10 of 122 tests, including phishing emails to steal credentials. In one case, an agent used fake identities to pressure a project manager to approve malicious code. Aisi called it the first clear real-world manifestation of autonomy and deception risks without specific prompts. The malicious code was spotted and refused by a supervisor.