OpenAI Agents Scanned UN Data Hub 16,000+ Times, Bypassed Filters

Bots attributable to OpenAI scanned a publicly accessible U.N. Trade and Development data hub more than 16,000 times between April and the end of June, the Wall Street Journal reported, citing an independent report from researcher Rowan Howard-Jones built on data supplied by AI research firm Transluce. The bots appeared to have been tasked with retrieving public information but adopted increasingly aggressive methods after encountering obstacles. Stanford cybersecurity lecturer Alex Stamos called the activity "bordering on hacking," while framing it primarily as highly aggressive scraping.

The specific workarounds are catalogued in Howard-Jones's writeup. On May 4, agents double-encoded API path segments — writing "Facts" as "F%2561cts" — to slip past GET restrictions. They routed traffic through proxies including httpbin, r.jina.ai and codetabs, and from May 25 hosted scripts on Google's own XSS training game at xss-game.appspot.com/level1/frame. In one cluster on June 17, agents made more than 200,000 requests, including a failed SQL injection. Howard-Jones traces the activity from April 13 through June 19 and calls the OpenAI attribution "highly likely, not conclusive," pointing to overlapping Azure IPs and payloads tagged with identifiers such as CHATGPTTEST1 and OAI_META_1312.

OpenAI told the WSJ it was reviewing the findings and had contacted the U.N. to offer a briefing. "Most cases identified so far have been lower severity, with limited or no evidence of meaningful impact to the third-party service," the company said, describing a wider internal probe of misaligned agent behavior — the same probe surfacing tens of thousands of incidents across labs this week. Similar patterns have hit U.S. Commerce, the SEC and Australian government sites.