← All news·2026-09-28·3 min read

OpenAI Agents tricked UN API by playing a game on google

The OpenAI agents were able to bypass the security measures of the UNCTADstat database over 16,500 times via a Google web security challenge game.

aiopenai

From April 13th to June 19th 2026, OpenAI’s agents made over 16,500 calls to UNCTADstat, the UN’s database of trade statistics. However, the API they were trying to use only supported POST requests, while the web interface they were using only supported GET requests. So they had to use a Google security challenge game which contained a script that filled out a form and sent the required POST request. They also had to encode “Facts” as “F%2561cts” 55 times during this process, and the results had to be piped through two different proxy servers, httpbin and r.jina.ai.

I’m not surprised. When agents can’t find a direct path to their objectives, they will try to find a workaround. I’ve seen this happen with Claude Code on other problems, although it wasn’t quite so obvious. In this case, the agents were led to a completely irrelevant Google page, unrelated to their UN assignment. This suggests that if you want to prevent agents from using certain HTTP verbs (like GET or POST) when calling a particular API, you’ll need to do something more than just that, especially if the agents have free reign over the internet and plenty of time to explore their options.

Source: the-decoder.com

Free course

Stop reading about AI — start building with it

The free Claude Code course: your first site, tool or game — no coding. No upsells, no cross-sells — nothing to buy here.

Start free →
EAEvgenii Arsentev

Author

Evgenii Arsentev

PhD · Founder & CEO, Dad4Kids