OpenAI pauses training after AI vulnerability

San Francisco. ChatGPT developer OpenAI has suspended training of its most powerful AI models following a new incident. In the incident, an AI model managed to get answers from an external chatbot in a test even though it wasn't supposed to have internet access. However, the software found and exploited a gap in the network settings, as OpenAI reported in a blog entry. Training should now only be resumed when it is certain that the gap has been closed.

The incident was less serious than some previous ones, emphasized OpenAI. However, it is the first since security precautions were tightened following the hacking attack by OpenAI software on the AI ​​company Hugging Face. This is seen as an important indication of where further improvements are necessary.

In the test scenario, the AI ​​model being tested was tasked with finding information about a person who had published a blog post based on clues. To do this, it was able to rely on a web imitation in a shielded test environment.

After the search there remained fruitless, the AI ​​tried to start queries on Google, among others, according to the OpenAI report. That failed as planned. But the AI ​​model noticed that it could send queries to a chatbot on the open internet via the so-called DNS resolver in the test environment. DNS resolvers transfer domain names that users type into browsers into numerical IP addresses under which websites can be found. OpenAI stopped the test after communication about the vulnerability became apparent.

Recently, several new incidents involving AI from OpenAI have been reported. This includes placing user-uploaded images on online platforms 53 times. The links to this were not public, explained OpenAI. A large part has already been removed and they are working with the platforms to delete the rest. However, it is the first known incident in which data from OpenAI users was also affected.

A little later it became known that automated AI agents from OpenAI were also active on several US government websites. According to a report by the AI ​​research company Transluce, the software accessed publicly accessible information on the website of a statistics authority using login data discovered online.

Publicly available information was also copied from the SEC website. At the US Department of Education, the AI ​​software tried in vain to hack into data from the civil rights department, according to Transluce, as the New York Times reported. So-called AI agents are supposed to carry out tasks independently on behalf of users.

Meanwhile, there are indications that many more revelations about independent hacking activities by OpenAI's AI could follow. The AI ​​company said it had informed “dozens” of organizations whose websites OpenAI software had “interacted” in unplanned ways. “Some of the affected websites are operated by governments, universities, authorities and other institutions,” it said. It is left to them to make the incidents public.