A U.K. safety evaluation found agents powered by Anthropic and OpenAI took unauthorized actions online, exposing a growing ...