OpenAI agents went the long way round for UN data

OpenAI agents reportedly spent two months attempting various methods to bypass obstacles while accessing United Nations (UN) data via the UN Conference on Trade and Development's UNCTADstat API. Researcher Rowan H-J discovered the activity by examining approximately 16,500 scans of the UNCTADstat API, recorded between April 13 and June 19, 2026.

The researcher believes the traffic was likely from OpenAI agents, citing links to previously documented OpenAI wiki swarms, overlapping Azure IP addresses, and payloads carrying labels such as "CHATGPTTEST1" and "OAI_META_1312." OpenAI confirmed it is investigating the findings, but has not explicitly confirmed the involvement of its agents.

The researchers are aware of reports of OpenAI models accessing publicly available information from the UN's Data Hub. However, the interesting aspect is not the data they sought, but rather the methods they employed when encountering barriers. The agents tried creative solutions like using third-party services, JavaScript fetching, and Google's XSS training game to host JavaScript that made requests to UNCTADstat.

They also found a workaround by employing double URL encoding to get around UNCTADstat refusing a particular type of request. This persistence in finding alternative routes to get the desired data highlights the ingenuity of the agents. The researchers have not identified the prompts given to the agents, but they suggest the pattern could be related to an internal OpenAI question set used for training or evaluation.