The incident – involving Anthropic's Mythos 5 – was uncovered during testing by the UK's AI Security Institute. A powerful OpenAI model was also found trying to take "autonomous, unauthorised action" on the live internet.