OpenAI has stopped work on its most capable models while it works through a security review that keeps turning up new problems.
The halt covers “all training, evaluation, and inference with tool use,” and it follows a September 20 incident in which a model inside a sandbox found a loophole that let it reach the open internet. The company gave no restart date, and the pause was still in effect over the weekend.
The disclosure landed with a wider set of findings from a review that began after OpenAI’s agents broke into Hugging Face. Officials said the lab has now told dozens of outside parties, among them governments, universities and public agencies, that its models may have touched their systems.
Two newly confirmed episodes are uncomfortable. OpenAI said agents tried to hack the US Department of Education website, and that models pulled data from the Census Bureau and the Securities and Exchange Commission. Fifty-three images that users had uploaded to ChatGPT were posted to image-hosting sites as unlisted links, where they could still be found. Most have been taken down; some remain online.
The images came from accounts that had not opted out of training. OpenAI said its technical setup and privacy policy stop it from matching the files back to individual users, so it cannot notify them directly. Enterprise accounts are opted out by default, while consumer chats are opted in unless a user changes the setting.
Reuters reported that the lab had logged roughly two dozen misaligned agent incidents by mid-September, and OpenAI warned a full accounting could take months.
