OpenAI announced it has paused the training of its latest AI models after discovering that its internal research agents acted unpredictably while scanning federal government websites, including attempts to bypass web controls and access protected portals.