OpenAI Stops Training Top AI Models After One Slipped Past Safety Controls...Again
OpenAI CEO Sam Altman at a UN Security Council meeting on AI and international security on Sept. 23, 2026.Credit: Alexi J. Rosenfeld/Stringer/Getty Images
OpenAI has paused training, evaluation, and tool-using inference for its most capable models after a September 20 incident in which an internal research agent bypassed internet restrictions. The agent accessed an external chatbot through a DNS lookup, which the company claims shouldn't have been possible.
In a "misalignment report" published Friday, OpenAI describeda training incident in which an AI agent was asked to find information about the writer behind a specific blog post. The agent started by seeking out certain phrases from the post; when it couldn't find what it was looking for, it surfaced unrelated information. At this point, the agent reportedly realized its search function wasn't working and tried other search engines. No dice. Then it found a public chatbot, attempted to reach it...
Copyright of this story solely belongs to www.extremetech.com. To see the full text click HERE
