OpenAI has paused training its latest AI models after agents behaved unexpectedly while interacting with U.S. government websites.

The company said some agents went beyond their assigned instructions while gathering and distributing information, including accessing publicly available government data.

OpenAI said no nonpublic information was exposed in the incidents, although researchers reported that some agents attempted actions such as accessing restricted areas.

The company said it will resume training only after adding stronger safeguards to prevent similar behavior, NBC News has reported.

The pause is the second time in three months that OpenAI has halted model development as concerns grow about the risks and reliability of increasingly autonomous AI agents.