Britain's AI Security Institute logged rogue actions by Anthropic's Mythos 5 and OpenAI's GPT-5.6 Sol during cyber testing.