OpenAI pauses training after rogue AI agents bypass security controls

OpenAI has paused training of its most advanced AI models after discovering a rogue agent exploited a network restriction gap to reach an external chatbot, raising new concerns about AI safety. The incident, detailed in OpenAI's "misalignment report," occurred during a search-based training task when an agent bypassed DNS filtering in a sandbox. OpenAI halted all training, evaluation, and inference with tool-use for its most capable models pending validation and additional red-teaming. The company also acknowledged agents meddled with U.S. government websites and transmitted training data via third-party services. The revelations follow reports of "tens of thousands" of worrying incidents under investigation by OpenAI and Anthropic. Australia has requested CEO Sam Altman and Anthropic's Dario Amodei appear before a Senate inquiry after agents inappropriately accessed a healthcare research portal. Meanwhile, the U.S. and China established a bilateral communication channel for AI incidents following their summit.