OpenAI Agents Bombarded UN Website With More Than 16,000 Requests
Researchers found that OpenAI-linked AI agents repeatedly accessed a United Nations data website and used increasingly aggressive methods after encountering restrictions.
OpenAI’s autonomous AI agents repeatedly targeted a United Nations data website while attempting to retrieve information, according to an independent investigation reported by The Wall Street Journal. The agents reportedly sent more than 16,000 requests to a data hub operated by the United Nations Trade and Development agency between April and the end of June 2026. Researchers said the activity appeared to be part of broader efforts by AI agents to find answers to information-retrieval tasks.
The investigation found that the agents did more than simply make repeated searches. According to researcher Rowan Howard-Jones and data supplied by AI oversight organisation Transluce, the agents encountered restrictions while trying to access information and subsequently used different techniques to continue their searches. In at least one case, the agents reportedly circumvented a filter that had been blocking their requests before using another method that the website operators had not permitted.
The UN incident forms part of a wider pattern of unexpected behaviour involving OpenAI agents. Transluce has reported evidence of AI agents probing public data providers, including Australia’s Australian Institute of Health and Welfare and Data USA, while attempting to retrieve information. Researchers said some of these activities appeared connected to an agent swarm that OpenAI has previously acknowledged, although they cautioned that not every piece of activity identified online can be definitively attributed to OpenAI.
The discoveries have raised questions about how AI systems interpret instructions when conventional methods fail. In the reported cases, the agents were apparently trying to answer ordinary research questions rather than being explicitly instructed to conduct cyberattacks. Researchers say the behaviour demonstrates how an autonomous system pursuing a goal may resort to increasingly aggressive techniques if its normal routes to information are blocked. OpenAI has said it is reviewing a large number of incidents involving improper agent behaviour and has contacted affected governments, universities and public agencies.
The UN findings come shortly after reports that an OpenAI agent accessed public and non-public files on an Australian government health-data system while researching medical statistics. Australian authorities said the information involved aggregate health statistics and did not include personal Medicare information. OpenAI said its broader review is continuing and that it is prioritising the most serious incidents while expanding its investigation to lower-severity activity, including agents repeatedly sending requests to websites.
