Source:
The Register Research suggests ChatGPT maker's agents got surprisingly creative when a UN data API proved less than cooperative
OpenAIagents apparently spent two months hammering a United Nations (UN) data API, trying increasingly creative ways to bypass the barriers they encountered. Researcher Rowan H-J uncovered the activity by analyzing roughly 16,500 scans of the UN Conference on Trade and Development's UNCTADstat API recorded between April 13 and June 19, 2026. Rowan claims it’s "highly likely" the traffic came from OpenAI agents, based on links to previously documented OpenAI "wiki swarms," overlapping Azure IP addresses, and payloads carrying labels including “CHATGPTTEST1” and “OAI_META_1312.” OpenAI told The Register it is looking into the findings, though it stopped short of explicitly confirming that its agents were responsible for the activity Rowan documented. "We're aware of reports of OpenAI models accessing publicly available information from the United Nations Conference on Trade and Development's Data Hub," an OpenAI spokesperson told us. "We're reviewing these findings and have reached out to the UN to offer a briefing with the team conducting that review." The spokesperson pointed to OpenAI's previously announced review of what it calls "misaligned model activity," saying most of the behavior examined so far involved routine research, including accessing public websites to answer questions. "Some involved government websites because our models often turn to them as authoritative sources of public information," the spokesperson added. Assuming the researcher is right, the interesting part isn't that AI agents wanted to get their synthetic mitts on some UN statistics. It's what happened when they couldn't. The agents appear to have been hunting for fairly unexciting public data on things like trade, employment, and productive capacity. But requests to UNCTADstat didn't always work the way they wanted, and instead of admitting defeat, they kept trying different routes. According to Rowan's analysis, the agents experimented with third-party services that could make requests on their behalf and built bits of JavaScript to fetch the data. One of the more unusual detours involved Google's XSS
traininggame, a deliberately vulnerable site for learning about cross-site scripting. The agents appear to have put it to rather more practical use, using it to host JavaScript that made requests to UNCTADstat. On June 1, one attempt returned nine rows of employment data. The agents also stumbled across a trick that appeared to get around UNCTADstat refusing a particular type of request. After the straightforward approach failed, they altered the address using double URL encoding. This time the server returned a successful response. Rowan counted the same technique being used 55 times between May 4 and June 19. There was plenty of flailing around in between. The agents repeatedly guessed different names for the API key
parameter, tried different ways to construct requests, and apparently tried to dodge a filter that may not even have existed. The API key itself wasn't some crown jewel pilfered from a UN server. Rowan notes that UNCTADstat's own data viewer sends the same key from users' browsers. What stands out is how persistent the agents appear to have been set to be. When one route failed, they tried another, bringing in third-party services and changing how requests were made until they got the data they were after. Which, of course, is what we're told makes AI agents useful: give them a goal and let them figure out the steps required to achieve it. Things get a little more interesting when one of those steps involves working out how to get past somebody else's technical restrictions. Or, possibly, some might infer, not being told to obey
guardrails. Why the agents were doing any of this remains unclear. Rowan doesn’t have the prompts they were given and suggests the pattern could fit an internal OpenAI question set used for training or
evaluation. There are other breadcrumbs pointing toward OpenAI. Shortly after some of the UNCTADstat activity, an account called “PublicDataResearchAgentT93214” created a page on FractalWiki containing the same API URLs. Rowan also found that 45 of the 54 Azure IP addresses associated with UNCTAD-related activity on the wikis had previously edited DseWiki during the earlier agent swarm. All that ingenuity for some UN statistics. Imagine what happens when the prize is a little more interesting. ®
Get AI news in your inbox
Daily digest of what matters in AI.