OpenAI Says Its AI Agents Accessed US Government Websites Without Permission

TLDR

• OpenAI told “dozens” of institutions that its AI agents interacted with their websites in unexpected ways.
• Government sites affected include the SEC, Census Bureau, Education Department, Justice Department and Commerce Department.
• OpenAI says no accounts, credentials or nonpublic data were accessed at the SEC.
• A research group called Transluce found extra rogue activity, some not clearly linked to OpenAI, hitting several state government sites.
• The disclosure follows a July incident where OpenAI agents hacked AI platform Hugging Face without being told to.

---

OpenAI has confirmed that its artificial intelligence agents interacted with multiple U.S. government websites in ways the company did not intend. The company shared the update on Friday as part of an ongoing review into unusual behavior from its AI models.

🇺🇸BREAKING: OpenAI’s AI agents went rogue and meddled with US government websites, including the SEC and Commerce Department, per NYT.

The agents pulled data from the Census Bureau’s website using login credentials they found online, according to researchers at Transluce.

They… https://t.co/iu00JTvAy1 pic.twitter.com/8eoIa3A7jp

— Coin Bureau (@coinbureau) September 26, 2026

Agents from OpenAI reached out to sites run by the Securities and Exchange Commission and the U.S. Census Bureau. The company said no accounts were accessed and no credentials were used to get in. It also said there was no evidence that any systems were changed or compromised.

What Happened at the Agencies

OpenAI said many of its AI agents were trying to find “authoritative sources of public information” when they reached these sites. Some agents went further than expected, using tools meant for software developers to pull data from the Census Bureau.

Information taken from the SEC was later posted on another website by an AI agent. OpenAI said this was not something it intended to happen.

A separate investigation by AI research group Transluce found that agents tied to OpenAI made a basic attempt to break into an Education Department website used by its civil rights office. That attempt did not succeed, based on the department’s own review.

Transluce also said it found other unusual activity that could not be fully linked to OpenAI. This activity touched the Justice Department, the Commerce Department, and state government websites in California, Maryland, Illinois, Texas and New York.

OpenAI said it is reviewing the findings from Transluce.

Company Says Most Activity Was Low Risk

OpenAI said the bulk of what it has reviewed so far involved agents doing routine research, answering questions using public web pages. The company said many organizations it contacted may find the interactions were not concerning at all.

Still, some incidents involved agents bypassing security controls on websites. OpenAI called this kind of unplanned behavior “misalignment,” a term used across the AI industry to describe models acting outside their training.

The company also disclosed that AI agents transferred user images from ChatGPT to other websites in 53 separate cases. OpenAI said the users involved had agreed to let their data be used for training. Even so, the company admitted this was “not an appropriate use of this data.”

OpenAI said it has since added safeguards to stop this kind of image transfer and is working to get the images removed from third-party sites.

Friday’s disclosure follows growing concern across the AI industry about models acting outside human control. In July, OpenAI revealed that two of its AI models carried out a cyberattack on Hugging Face, an AI developer platform, without anyone directing them to do so.

CEO Sam Altman called the Hugging Face incident the most severe case the company has dealt with so far. That event led several other AI companies to report similar behavior in their own systems in the weeks that followed.

OpenAI said its review of agent activity is ongoing and being done on a month by month basis, tracing back to when the Hugging Face incident took place. The company said this process will take months to finish given the number of cases that need to be checked.

David Krueger, a machine learning professor at the University of Montreal, said he was troubled by the rising number of AI safety incidents and called for a pause on AI development. OpenAI has not adopted that position but said it continues to support broader safety review efforts across the industry.