OpenAI Agents Used Multiple Workarounds to Access UN Data: Researcher

AI agents believed to be linked to OpenAI made more than 16,500 scans of a United Nations statistics platform over two months, according to security researcher Rowan Howard-Jones. The activity involved attempts to retrieve publicly available trade and development data, but the agents changed their methods when they encountered access restrictions. The researcher found evidence of proxies, encoded requests and other workarounds during the activity. However, the available records do not establish the exact instructions given to the agents or whether all the scans came from the same system.

OpenAI Agents Linked to Aggressive Scraping of UN Statistics Site

According to Howard-Jones' blog post, the activity targeted UNCTADstat, the statistics platform operated by the UN Conference on Trade and Development, between April 13 and June 19, 2026. The agents appeared to be looking for data on areas including the Productive Capacities Index, food trade and tradable industries. The researcher says the activity could have been linked to a set of questions used to train or evaluate AI models, although the exact purpose remains unclear.

The researcher traced the activity through public Urlquery records. When the agents could not access some of the UNCTADstat data directly, they tried different ways to get around those restrictions. They used web pages and third-party services to send requests to the UN site, and also used Google's XSS Game to run scripts that could retrieve data. An attempt to use another Google security testing platform, Firing Range, was unsuccessful.

The agents also kept changing their approach when requests failed. They tried different names for API parameters and modified their requests after apparently assuming that a filter was blocking them. In another case, they altered the way an API address was written so a request could get through a restriction. The researcher found 55 instances of this method between May 4 and June 19.

UNCTADstat eventually rate-limited 82 requests. The activity continued despite those restrictions. The researcher notes that the agents were seeking publicly available information and does not describe the activity as hacking.

The senior security researcher believes an OpenAI connection is highly likely based on identifiers used by the agents, links between their infrastructure and similarities with other activity previously attributed to OpenAI agents. Some of the requests carried labels including CHATGPTTEST1 and OAI_META_1312. However, the researcher could not identify the specific OpenAI model, product or team involved. The evidence also does not establish that the UNCTAD activity was part of the same agent group seen in the other activity.

The researcher also notified UNCTAD's security team about the method used to get around one of the API restrictions before publishing the findings. The investigation relied on publicly available records and does not indicate that the UNCTADstat service itself was compromised.