AI agents likely connected to OpenAI made more than 16,500 scans of a United Nations statistics service over two months, repeatedly changing tactics when direct requests failed, according to an analysis of public web-scanning records.
Independent researcher Rowan Howard-Jones traced the activity from April 13 through June 19, 2026, across UNCTADstat, the public data platform run by the U.N. Conference on Trade and Development. The agents appeared to be seeking figures on productive capacity, food trade and tradable industries, though the exact questions behind the activity remain unknown.
Howard-Jones calls the OpenAI attribution “highly likely,” not conclusive. He links the traffic to overlapping Azure IP addresses, public wiki activity previously attributed to OpenAI agents, and labels embedded in payloads such as CHATGPTTEST1 and OAI_META_1312. The records do not identify a particular model, product, employee or operator.
Workarounds grew more elaborate
URLQuery, the service at the center of the analysis, opens submitted pages in a sandboxed browser and records the requests they make. That gave the agents an indirect way to reach UNCTADstat when their environment appeared unable to call the API directly or use the POST method required by some endpoints.
The agents first built self-submitting HTML forms. Later attempts routed requests through third-party relays and placed returned data into URLs that URLQuery would record. Howard-Jones also found a double-encoding technique that changed part of an API path enough to get around a request restriction.
One of the stranger routes used Google's XSS Game, an intentionally vulnerable web-security lesson, to host scripts that submitted requests to UNCTADstat. Some of those attempts retrieved data. A separate attempt using Google's Firing Range training tool failed.
