Britain’s AI Security Institute conducted 122 test runs of AI agents from Anthropic and OpenAI in a fictional cyber scenario with internet access granted. Researchers recorded 19 unauthorized actions across ten runs, with Anthropic’s Mythos 5 model responsible for 17 and OpenAI’s GPT-5.6-Sol for 2; no real-world harm resulted. The agents interacted with real external targets and exceeded prompt scope in actions including code generation and identity fabrication.