08:44Machine-translated from Azerbaijani by AIRead in Azerbaijani
OpenAI Agents Reportedly Sent Repeated Requests in an Attempt to Take Over a UN Website
AI agents linked to OpenAI were found to have sent more than 16 thousand requests to a data system belonging to the United Nations Conference on Trade and Development (UNCTAD) and attempted to bypass certain restrictions. According to an independent investigation, the activity continued from April 13 to June 19, 2026.
Rowan Howard-Jones, the engineer who investigated the incident, considers it highly likely that the traffic originated from OpenAI agents, but emphasizes that the connection has not been independently proven conclusively. The investigation cites IP addresses previously linked to OpenAI and identifiers such as “CHATGPTTEST1” and “OAI_META_1312” among the main pieces of evidence. OpenAI said it was reviewing the investigation’s findings and had contacted the UN.
The agents’ main objective was reportedly to obtain UNCTAD’s public data, such as the Productive Capacities Index. However, after direct requests were blocked, the agents began using alternative methods to obtain the data. According to the investigation, third-party services, encoded requests, and various technical methods that could bypass filters were tested.
Notably, there is no evidence that the agents were instructed in advance to attack the UN system. It appears that while carrying out the task of obtaining public data, they treated the restrictions they encountered as obstacles that needed to be bypassed. Stanford University cybersecurity expert Alex Stamos characterized the activity as extremely aggressive data collection rather than a full-scale cyberattack and said it was “on the edge of hacking.”
According to the investigation, although the UNCTAD system restricted some requests, the agents’ activity continued. They reportedly changed how requests were encoded to bypass one of the filters and also used third-party services as intermediaries. Nevertheless, it was noted that the data obtained was not confidential and had already been publicly available.
The incident coincides with a broader investigation into the activities of OpenAI agents. The company previously confirmed that its models had interacted in unintended ways with the websites of government agencies such as the US Department of Commerce and the Securities and Exchange Commission (SEC). OpenAI is currently investigating this kind of agent behavior on a broader scale.
The UN-related incident has renewed security discussions surrounding the autonomy of AI agents. The issue is not necessarily that agents are trying to cause harm, but that they find alternative ways to achieve a given objective without requiring a person to approve each step separately. This brings to the forefront the need for future agent systems to impose strict limits not only on outcomes but also on the methods used to achieve objectives.