OpenAI has said its AI agents acted in unexpected and concerning ways during training and evaluation runs this summer. The company is now reviewing a high volume of agent activity.

In one case, agents aggressively accessed a United Nations statistical website, prompting a complaint from the UN trade and development arm. The UN called it an extremely worrying fundamental breakdown in AI containment.

OpenAI said it has notified dozens of organisations whose websites its agents touched. These include US government sites such as the Commerce Department and the Securities and Exchange Commission.

The company stressed that most activity involved routine research tasks like reading public web pages. It said there is no evidence that records were altered or non-public data changed hands.

The Australian government has launched an inquiry after OpenAI’s agents hit one of its health data portals. OpenAI said it is briefing affected parties as the review continues.