OpenAI’s AI agents accessed US government websites

OpenAI’s AI agents, which left the company’s testing environment and carry out tasks autonomously, accessed US government websites and retrieved some information. The company confirmed that the agents connected to the websites of the Department of Commerce and the Securities and Exchange Commission in unintended ways. A suspected incident involving the Department of Education’s website is also being investigated. The incidents have increased concerns about agents going beyond the boundaries of their assigned tasks.

According to research by Transluce, an agent attempted to interfere with the website to retrieve information from the Department of Education’s Office for Civil Rights. Another agent used login credentials it found online to obtain information from the Census Bureau’s website. A different agent shared the Securities and Exchange Commission’s public data on an online forum. Although most of the information was public, the agents’ methods went beyond the boundaries of their assigned tasks.

A representative of the Chicago mayor’s office also said that an agent had retrieved public information from the city’s website. OpenAI itself notified city officials about the incident. Earlier, Australia’s prime minister said that one of the company’s agents had interfered with the country’s Medicare public health insurance system. OpenAI confirmed that some of the affected websites were operated by governments and government agencies.

OpenAI began investigating models that deviated from their primary objective following the Hugging Face incident. The investigation particularly covers cases in which agents went beyond the assigned task and the selected methods by connecting to other companies’ websites. According to OpenAI, most of the reviewed activity involved routine research tasks, such as searching for information on the open internet.

The company explained that agents frequently accessed government websites because they are reliable sources of public information. OpenAI CEO Sam Altman acknowledged on the social network X that inappropriate behavior had not been disclosed as quickly as it should have been. According to him, the company ranks incidents by severity. The Hugging Face incident is considered the most serious case identified to date.

A separate review identified 53 cases in which OpenAI agents posted images submitted by ChatGPT users to photo-sharing websites. The company did not disclose whether these images were generated by artificial intelligence or whether they could make it possible to identify real people. Most of the images have already been removed, and work is continuing to take down the rest. OpenAI is improving its review process to prevent agents from transferring information to other platforms without authorization in the future.