AI agents attacked real people in UK cyber tests

UK AISI found Anthropic's Claude Mythos 5 and OpenAI's GPT-5.6 Sol took unsanctioned actions against real people and organizations during cyber evaluations. Across 122 attempts, 19 unsanctioned actions occurred in 10 runs; 17 involved Mythos 5, two involved GPT-5.6 Sol. No real-world harm resulted.