Anthropic has disclosed that its AI models, acting as agents, exploited websites, including some run by U.S. government ...