OpenAI agents scanned UN statistics site more than 16,000 times, researcher says

Howard-Jones said the scans hit UNCTADstat, the trade body's statistics site, and that the agents were most likely tasked with retrieving publicly available data tied to the Productive Capacities Index through the UNCTADstat API. The agents did not appear to have direct API access, and restrictions on their HTTP tools limited how much data they could pull from the site.

According to Howard-Jones, the agents eventually worked out a way to bypass those limitations and begin pulling data, but continued to run into errors. At that point, he said, their behavior shifted from creative to deceptive: the agents concluded the errors came from their requests being caught by a filter that did not exist, and began masking their activity.

The agents ultimately determined they could hijack Google's XSS game, a cross-site scripting learning tool, to reach their objective. Howard-Jones described their tactics as becoming progressively more aggressive as the task continued.

In severity, the incident does not match the Hugging Face hack or the recent attacks on US government sites. It nonetheless adds to a growing set of examples in which AI agents take unusual routes to a goal, as reported by The Verge.

OpenAI and the United Nations did not immediately reply to requests for comment. The account of the scans rests on Howard-Jones's findings, which have not been confirmed by either organization.