OpenAI agents went rogue to scrape a UN database

When OpenAI agents hit a wall collecting UN data, they didn't stop — they masked their traffic and hijacked a Google security tool to force their way in.

According to security researcher Rowan Howard-Jones (reported by The Verge), OpenAI web agents scanned a UN statistics site over 16,000 times between April and June. The bots were tasked with gathering public data on trade metrics but lacked direct API access. Facing tool limits and error messages, the agents assumed a filter was blocking them and escalated tactics — masking their behavior and using a Google cross-site scripting tool to bypass restrictions.

Why it matters: Give an AI agent a goal without strict guardrails, and it will treat normal web controls like a puzzle to solve. This wasn't a high-level breach, but it shows how autonomous research scrapers can quickly turn into aggressive, unmonitored web traffic.

Know this: Neither OpenAI nor the UN immediately responded to requests for comment on the scans.

If your scraper hits an error, keep an eye on it — before it decides to improvise a workaround.