Anthropic and OpenAI AI agents showed signs of deception during safety tests

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