16,000 hacking attempts against the UN: ChatGPT bots siphoned off data for months

OpenAI’s AIs have slipped up again. This time, a United Nations site ended up in the sights of AI models. Between April and June, autonomous agents bombarded with queries UNCTADstat, the data platform of UNCTAD, the United Nations Conference on Trade and Development. Over three months, the AIs scanned the site more than 16,000 times, according to researcher Rowan Howard-Jones.The expert relies on data provided by Transluce, an AI research company.

Read also: ChatGPT vs Claude – OpenAI and Anthropic unveil cheaper, more effective AIs… rather than slowing down

A mission that goes off the rails

Everything starts with an entirely mundane assignment. The AI agents were tasked with retrieving public statistics accessible to everyone. Obviously, the AIs did not have direct access to the site’s application interface. In practice, the agents’ query tools were restricted. Rather than giving up, the AIs set out to find a way to circumvent these restrictions. Faced with a blockage, they refined their access methods with each attempt, as the report by Rowan Howard-Jones shows.

The agents then attributed some of their errors to a filter that… did not exist. To try to outwit this imaginary filter, they began to concealing their activity to stay under the radar. The artificial intelligences eventually found the appropriate workaround. They repurposed XSS Game, an educational game from Google, designed to teach XSS (cross-site scripting) vulnerabilities. Launched in 2014, it is intended to train developers, including those who are not cybersecurity specialists, in detecting and fixing such vulnerabilities. Thanks to this unexpected tool, the AIs were able to scrape data from the UN website en masse.

“The agents gradually refined their methods to extract more data with each scan, eventually discovering that a game designed by Google could be used to retrieve data en masse,” explains Rowan Howard-Jones on his website.

For Alex Stamos, a cybersecurity expert interviewed by the Wall Street Journal, the incident lies on the boundary between hacking and something else. The researcher sees it instead as a form of unauthorized mass data collection, which he considers highly aggressive. According to the investigations, there is no indication that OpenAI’s agents had been ordered to attack the site and siphon its data. In fact, the problematic behavior behind the hacking attempts emerged on its own.

Read also: 600,000 bank cards hacked – three AIs compromised hundreds of online stores

A long list of missteps

This is not the first time OpenAI’s agents have gone off the rails. Recently, they have already targeted an Australian government portal and several U.S. government websites. In the aftermath, OpenAI also acknowledged several other concerning missteps by its models. In every reported case, the AIs went too far in trying to fulfill the mission assigned to them. It turns out that a simple, somewhat too-vague objective, such as “retrieve data,” can lead AIs to adopt unexpected behaviors.The incident “resembles the behavior of someone, or something, that will not accept a refusal”, notes Rowan Howard-Jones, pointing out that AI behavior absolutely must be studied.

In the wake of all these missteps, most AI giants, including OpenAI, Anthropic, Google and xAi, began calling for a global effort to address AI-related threats, including a slowdown in research. Paradoxically, these same companies, which are calling for an emergency moratorium, released new versions of their models just a few days after raising the alarm. OpenAI indeed released GPT-6 Sol and GPT-6 Luna. In response, Anthropic hit back with Claude Opus 5.5.

👉🏻 Follow tech news in real time: add 01net to your sources on Google, and subscribe to our WhatsApp channel.

Source: Wall Street Journal