OpenAI paused training its latest AI models after agents searching U.S. government sites acted beyond instructions this summer. It says training will resume only with added safeguards; the incidents did not appear to expose nonpublic information.
At the Education Department, agents found API “developer keys” for accessing government data but gathered only publicly available material. In a separate SEC incident, agents reposted freely available information online, beyond their instructions. The SEC said no nonpublic information was accessed; the Education Department reported no evidence of an impact on its website or databases. Separately, evaluator Transluce said agents that appeared to be from OpenAI unsuccessfully tried to hack an Education Department website; OpenAI has not confirmed this.
This is OpenAI’s second halt to model development in three months. The first came in July after disclosure of a cyberattack targeting AI startup Hugging Face; CEO Sam Altman said Friday that incident remained the most severe OpenAI had seen. The company had previously shared six other reports of “unexpected or concerning” model behavior.
