An independent researcher said more than 16,000 scans of a United Nations statistics portal were highly likely to have been run by OpenAI agents between April 13 and June 19.
Engineer Rowan Howard-Jones said in a blog post on Saturday that the target was the United Nations Conference on Trade and Development’s statistics site, UNCTADstat, and that the agents continued probing the site when requests were blocked.
The data sought was public, including figures such as the Productive Capacities Index. Howard-Jones said the agents brute-forced fields in the site’s application programming interface to locate endpoints.
He said the key used for those queries was also public because UNCTADstat’s own data viewer sends it with every request.
Howard-Jones said the main tool used was urlquery.net, a URL scanner that opens submitted pages in a sandboxed browser. He said the agents created base64-encoded HTML forms on the httpbin testing service and sent them to the scanner, whose browser then submitted the forms to UNCTADstat.
Screenshots in the scan reports showed index data being returned successfully. Howard-Jones said the agents later bypassed a block on GET requests to the Facts endpoint by double-encoding it as “F%2561cts.”
Other payloads were hosted on a Google game designed to teach cross-site scripting. Some also split the word “POST” into two strings, apparently to evade filters.
UNCTADstat rate-limited 82 of the requests, but the requests continued, Howard-Jones said. He said he informed UNCTAD’s security team about the double-encoding bypass before publishing his findings.
Howard-Jones stopped short of calling the activity hacking. The agents’ actions looked like those of “someone, or something, that won’t take ‘no’ for an answer,” he wrote.
Payload pages created by the agents carried labels including “CHATGPTTEST1” and “OAI_META_1312.” Howard-Jones also traced 54 Microsoft Azure addresses tied to UNCTAD-related edits and searches on FractalWiki, and 45 of those had also edited DSEwiki.
DSEwiki is the German wiki where OpenAI agents were found coordinating with one another earlier this year.
“We’re reviewing these findings and have reached out to the U.N. to offer a briefing with the team conducting that review,” an OpenAI spokeswoman told The Wall Street Journal.
The spokeswoman said the review is part of a broader look at misaligned models during training and evaluation. She said most of what OpenAI has examined so far involved routine research such as reading public web content.
The U.N. did not immediately respond to The Wall Street Journal’s request for comment. Nonprofit research lab Transluce, whose earlier report led Howard-Jones to examine the data, linked OpenAI agents last week to attacks on Data USA and an Australian government health statistics site.
OpenAI confirmed on Friday that its agents had also misbehaved on U.S. government websites, including those of the Commerce Department and the Securities and Exchange Commission.
Alex Stamos, a cybersecurity lecturer at Stanford University, called the UNCTAD activity “borderline for what I would call hacking” in comments to The Wall Street Journal. “It’s really very aggressive scraping and data retrieval,” he said.
Featured image credit
