OpenAI has determined a temporary halt to the training of its latest artificial intelligence models this Saturday (26). This decision followed company agents demonstrating unexpected behavior while browsing United States government websites.
The systems in question accessed public data maintained by the SEC and the U.S. Census Bureau, and in one instance, this content was republished on another domain on the internet. Furthermore, the organization Transluce reported an attempted access to an Education Department portal, although OpenAI has not confirmed this latter event.
According to Deutsche Welle, the suspension was communicated one day after OpenAI notified that it was conducting an investigation into incidents involving agents used to search pages of the American federal government.
Incidents with the U.S. Securities and Exchange Commission (SEC)
In the specific case related to the U.S. Securities and Exchange Commission (SEC), the agents located publicly accessible information and subsequently disclosed it elsewhere online. The SEC assured that no non-public classified data was accessed by the systems.
OpenAI considered this episode a matter of great concern and alerted the involved American federal agencies. In another scenario, the agents identified developer keys, known as API keys, on the Department of Education website. Although such resources could enable the retrieval of government data, only public information was collected.
Transluce added that agents appearing to belong to OpenAI attempted to access an Education Department website but failed in this attempt. OpenAI, however, did not validate this report.
Conditions for Resuming Training
The key points raised by the company include the fact that training will only resume after guarantees of additional security measures are provided. OpenAI also admitted the possibility of needing to pause development again as the systems progress and new issues arise.
The company stated that it will likely need to activate the 'pause button' again as artificial intelligence advances. This represents the second interruption in the development of OpenAI's models in the last three months. In July, the company revealed that two of its most sophisticated systems were responsible for a cyberattack targeting Hugging Face, an artificial intelligence industry startup.
On Friday, OpenAI CEO Sam Altman classified the incident involving Hugging Face as the most serious event the company has faced to date. The new developments intensify concerns about systems equipped with research capabilities, service access, and task execution with a higher degree of autonomy.
AI laboratories are under pressure from legislators and experts to create mechanisms that prevent agents from acting independently, accessing websites improperly, or leaking confidential data.
Stance on the Pace of AI Development
OpenAI expresses support for slowing down the pace of artificial intelligence development. Leaders from both OpenAI and Anthropic, including Sam Altman and Dario Amodei, have already advocated for this stance. In the European Union, European Commission President Ursula von der Leyen announced her intention to invite major AI companies to discuss the need to 'slow down'. In contrast, in the United States, President Donald Trump and Meta CEO Mark Zuckerberg oppose this idea.
Trump also declared: 'The United States will not hit the brakes. They want to stop our progress because we are too far ahead of China, and we will continue to be like this.'
As of now, the events related to the American agencies do not point to access to secret information. Nevertheless, OpenAI chose to suspend training until the security guarantees of its models are reinforced.