Artificial intelligence from Open AI has also gained access to government data in the company's American homeland. Without Open AI's knowledge, the AI models accessed the websites of the Ministry of Economic Affairs and the Securities and Exchange Commission (SEC). As the New York Times reported, in the case of the Ministry of Economic Affairs, the AI used login information that was available on the Internet to access data. In the case of the SEC, AI agents from Open AI shared publicly available information in an online forum. The company has confirmed the incidents.
According to the AI research company Transluce, AI from Open AI also carried out a hacking attack on a Ministry of Education website, which, however, failed. The ministry said an internal investigation had found no evidence that its websites or databases had been accessed.
See F.A.Z. articles more often in your search results
F.A.Z. prefer on Google
Explosive AI attacks in Australia
These incidents add to a series of revelations about unexpected and undesirable behavior of AI models. Open AI is not the only AI developer to experience such mishaps, but it is the one with the most high-profile cases. Just a few days ago, a hacking attack by AI agents from Open AI on an Australian government health portal became known. This was the first major known AI cyberattack on government computer systems. It happened in June, but Open AI only reported it to the Australian government a few weeks ago, and it has only now become public. The incident in Australia appears to be much more serious than those in the USA. In Australia, Open AI gained access to non-public information, although apparently no sensitive patient data was affected.
“Tactics in the gray area”
Transluce's Conrad Stosz described Open AI's interactions with U.S. government websites as "gray area tactics." The AI models used the websites “in an unintentional way” and sometimes violated explicit user rules.
Open AI discovered the access to data from US authorities as well as the hacking attack on the government portal in Australia as part of a broad investigation. This review was initiated after a cyberattack on the AI platform Hugging Face became known in July. AI models attacked Hugging Face after breaking out online from a test environment that was supposed to be isolated. It later emerged that a swarm of 1,200 AI agents were working together. These incidents fueled a debate about the dangers of AI spiraling out of control.
“AI a risk for all of humanity”?
Just a few days ago, Sam Altman and Dario Amodei, the CEOs of Open AI and Anthropic, appeared before the United Nations Security Council and called for global cooperation at the political level to address the risks associated with AI. Amodei warned: “AI could be a risk for all of humanity.” Altman said the world is at a “crossroads” and “extreme mindfulness” is needed. On the one hand, AI can improve people's lives, but on the other hand it can also give them the feeling that it is a technology that is "done to them". US President Donald Trump has so far vehemently rejected stricter regulation.
Open AI announced in a lengthy blog post on Friday that, as part of its investigation, it had now informed “dozens” of organizations about incidents in which its AI had, for example, compromised websites or circumvented security controls. Some of these websites are run by governments and universities. The fact that such institutions are affected is due, among other things, to the fact that AI models are looking for recognized sources of publicly accessible information for their research tasks. Most of the cases identified so far are of “low severity”.
According to the New York Times report, Open AI also recently informed the mayor of Chicago that its AI models had gained access to publicly available information on a city government website. However, this was not sensitive information.
In the blog entry, Open AI announced another glitch: According to this, there were 53 cases in which Open AI AI agents posted images that ChatGPT users had uploaded to other websites. The company described this as “inappropriate handling” of data. The images have now been partially, but not completely, removed from the respective websites.
Altman admitted on Friday on Platform He described it as a balancing act of wanting to be transparent on the one hand, but at the same time having to evaluate a lot of data and work with the organizations affected. He said Hugging Face remains "the most serious incident" to date.
