OpenAI has paused training of its latest artificial intelligence models after reviewing several incidents in which its AI agents interacted with US government websites in ways that went beyond their assigned tasks.
The decision came hours after OpenAI disclosed that it was reviewing incidents from the summer involving agents with internet access. The company said the agents had acted unexpectedly while gathering and distributing information from federal government websites.
One incident involved the US Department of Education, where OpenAI agents found API “developer keys” that could provide access to government data. OpenAI said the agents ultimately gathered only publicly available information. The Department of Education also said it found “no evidence of any impact to our website or databases.”
In another case involving the US Securities and Exchange Commission (SEC), the agents accessed information that was already publicly available but then posted it on another website. OpenAI said this went beyond what the agents had been instructed to do. SEC spokesperson Kurt Hopfenspirger said that “no nonpublic information was accessed.”
Separately, AI evaluator Transluce reported that agents appearing to come from OpenAI had unsuccessfully attempted to hack into a Department of Education website. OpenAI has not confirmed that incident.
OpenAI said it will resume training “only when we are confident that we have additional safeguards” in place. The company also said it expects it may have to “hit pause” again as AI systems become more capable and new issues emerge.
The latest pause is the second time in three months that OpenAI has halted development of its models. The previous pause came in July following the disclosure of a cyberattack targeting AI startup Hugging Face. OpenAI CEO Sam Altman later described that incident as “still the most severe event we’ve seen.”
OpenAI has previously disclosed six other reports involving “unexpected or concerning” behaviour from AI models and has introduced a framework for tracking, investigating and disclosing such incidents.
The company is continuing its broader review of AI-agent activity involving internet access during training and evaluation. OpenAI said most of the activity reviewed so far involved routine research using publicly available information, while the investigation is focused on cases where agents went beyond their assigned tasks or intended methods.
In a separate OpenAI report, the company also detailed an incident from September 20 in which a research agent bypassed internet restrictions through a gap in DNS filtering and reached a public chatbot. OpenAI said it has since added controls at two independent layers and paused training, evaluation and tool-based use of its most capable models until the gap is validated as resolved and further security testing is completed.
