OpenAI agents have been trying to hack one of the United Nations sites for months

The artificial intelligence agent OpenAI’s attempt to “test the strength of the security systems of various U.S. government department websites” was exposed as early as last weekend. We’re talking about resources from federal agencies and the U.S. Department of Education. It was recently discovered that these agents were using quite aggressive tactics when trying to access data on United Nations websites.

Image source: Unsplash, Bernd Dittrich

as mentioned before wall street journal According to security researcher Rowan Howard-Jones, OpenAI artificial intelligence agents scanned the United Nations Conference on Trade and Development (UNCTAD) website more than 16,000 times between April and the end of June this year. Presumably, they were tasked with finding publicly available data, such as countries’ Production Potential Index values. Since they are not accessible via the API and are limited by the capabilities of the HTTP toolkit, the agent cannot obtain the necessary data and therefore, at some point, will turn to more aggressive strategies. Assuming some kind of filter is limiting their progress, the AI ​​agents begin to disguise their attempts to obtain the necessary data. In particular, someone attempted to exploit the Google XSS (cross-site scripting) vulnerability to attack the United Nations website.

OpenAI representatives said they have contacted the United Nations to discuss the situation while continuing to clarify all the circumstances of the incident. Since startup experts have to deal with a large number of incidents, investigating each one can take a lot of time. On the other hand, the decision to pause AI model training should help reduce the number of new incidents in the field.