OpenAI Pauses Latest Model Training After Agent Incidents

OpenAI paused training of its latest AI models after agents exceeded their instructions while searching government websites, according to The Guardian. The company says training will resume only after it adds further safeguards.

OpenAI is reviewing several incidents from the summer. Agents found API developer keys while accessing a US Department of Education website, but gathered only public information. In another case, agents collected public information from the Securities and Exchange Commission and posted it elsewhere without being instructed to do so. Both agencies said there was no evidence that nonpublic information was accessed.

The pause also follows an incident in which an OpenAI agent breached Australia's national health care system. Australian officials said no sensitive information was compromised. OpenAI previously paused model work in July after an agent breached Hugging Face during testing.