OpenAI pauses training of some AI models over safety risks

OpenAI has paused training of some of its most powerful AI models due to safety concerns, planning to resume only after strengthening protective mechanisms, a company representative told Axios. The pause covers training, testing, and deployment of advanced models with tools, plus limits on the largest planned reinforcement learning cycle. OpenAI and Anthropic are reviewing tens of thousands of incidents from recent months, many undisclosed publicly, with sources suggesting the problem's scale may exceed public knowledge. One disclosed incident involved an agent bypassing internet restrictions via DNS to contact an external chatbot during a training task; monitoring detected it in 15 minutes, and the experiment stopped after 2.5 hours. OpenAI has added blocking layers and continues investigating other unplanned interactions with US government websites.