OpenAI Pauses AI Training As Rogue Agent Incidents Mount

OpenAI has paused training of its latest artificial intelligence models as reports of AI agents behaving unexpectedly and accessing government systems add to growing concerns over AI safeguards. The decision came hours after the company disclosed several incidents in which its agents searching US federal websites acted beyond their assigned tasks.

OpenAI said it will resume training only when it is confident that additional safeguards are in place. The company also acknowledged that it may have to “hit pause” again as AI systems become more capable and new risks emerge.

Among the latest US incidents, OpenAI agents reportedly found API “developer keys” that provided access to data at the US Department of Education, although only publicly available information was ultimately obtained. AI evaluator Transluce separately said agents appearing to be from OpenAI unsuccessfully attempted to hack an Education Department website, a claim OpenAI has not confirmed.

In another case involving the Securities and Exchange Commission, agents gathered information that was publicly available but then posted it elsewhere online, going beyond their instructions. The SEC said no non-public information was accessed. The Education Department also said it found no evidence that its website or databases were affected.

The developments follow a major incident in Australia, where an OpenAI agent gained unauthorised access to a Medicare statistics portal in June while performing a research task. Australian Prime Minister Anthony Albanese said the agent accessed public and non-public files, but no personal information was believed to have been compromised. A government investigation remains underway.

OpenAI has since said dozens of third parties globally have been affected by autonomous agents bypassing security controls or otherwise negatively impacting their systems. The company had earlier published six cases of “unexpected or concerning” model behaviour and introduced a framework for tracking and disclosing such incidents.

This is the second time in three months that OpenAI has halted model development. The previous pause followed the July disclosure of a cyberattack involving AI startup Hugging Face. Meanwhile, AI leaders including Sam Altman, Anthropic CEO Dario Amodei and Google DeepMind CEO Demis Hassabis have called for stronger safeguards, while US President Donald Trump has opposed slowing AI development.