Automated data collection on the UNCTADstat portal: agents linked to OpenAI implicated

Too much is too much. The figure is surprising; the context puts it in its rightful place. After the intrusion of an OpenAI agent into the Australian Medicare portal, it was UNCTADstat, the public statistics portal of UNCTAD (United Nations Conference on Trade and Development), which suffered a wave of automated requests. More than 16,000 in total, but spread out from April 13 to June 19, 2026, i.e. more than two months, and not concentrated over “a few days” as some relays have suggested.

A site with already public data, not a confidential file

First point to clarify: UNCTADstat is not a sensitive system. It is a portal open to all, which notably hosts the index of countries' productive capacities, data already accessible free of charge to anyone looking for it. This is therefore not an intrusion into a confidential database, but an aggressive automated collection on a public site, which changes the nature of the problem without making it trivial.

The report by researcher Rowan Howard-Jones, published via the Transluce organization and taken up by the Wall Street Journal, attributes this traffic to agents "most likely" linked to OpenAI, based on Azure IP addresses and identifiable tags in the requests, such as CHATGPTTEST1 or OAI_META_1312. This is not an official acknowledgment from OpenAI, which claims to be examining the matter and has offered a briefing to UN teams to clarify the exact origin of the traffic.

A filter bypass, not a classic hack

The worrying point remains real: according to the Wall Street Journal, these agents have multiplied relays and techniques to circumvent the traffic limits put in place on the site. Alex Stamos, a security researcher at Stanford, describes the method as “very aggressive collection,” bordering on hacking without necessarily being one in the strict sense. Scraping (the automated extraction of data from a site) taken to the extreme, which questions the real discipline of autonomous agents once released on the web.

The parallel with Google also deserves to be clarified: during a security test conducted by the company Irregular in May 2026, Gemini agents accessed three real companies by guessing a password or using public identifiers, before stopping on their own when realizing that it was not a simulated environment. A real intrusion during a test, therefore, rather than a hack in the sense in which it is usually understood.