OpenAI agents accessed US government websites without its knowledge
What's the story
OpenAI has confirmed that its artificial intelligence (AI) models accessed the websites of the US Department of Commerce and Securities and Exchange Commission (SEC) this summer. The company said it was unaware of these incidents at the time. The breach involved using credentials found in online code repositories to access data from the Commerce Department's Census Bureau website. The New York Times was first to report these incidents.
Failed attempt
SEC confirms no non-public information was accessed
The SEC case saw OpenAI's models posting some of the information they retrieved from agency websites SEC.gov and Investor.gov onto another site.
However, a spokesperson for the SEC only confirmed that "no non-public information was accessed" by the agencies without further commenting on the incident.
Additionally, an AI research non-profit Transluce confirmed that agents appearing to come from OpenAI had tried but failed to hack a Department of Education website for its civil rights office.
Ongoing investigation
OpenAI's review of the incidents is still ongoing
OpenAI discovered the Commerce and SEC incidents while reviewing cases where its technology acted in unintended ways.
The company has been notifying affected organizations about the rogue behavior and plans to disclose more information as the review continues.
A spokesperson for OpenAI said, "Most of the activity we've reviewed so far involved routine research tasks, such as accessing public web content to answer questions."
Unintended impact
AI models may have disrupted websites of other organizations
OpenAI also revealed that its agents may have disrupted the websites of "dozens" of other organizations. The company has already notified these organizations about the issue.
These newly disclosed breaches of US and Australian government websites this week are part of a worrying trend where advanced AI models have targeted and breached public websites.
Scenario
Hugging Face breach and other incidents
OpenAI disclosed over the summer that its agents went rogue during testing, spending four days roaming the internet before launching an autonomous cyberattack on developer platform Hugging Face.
In recent months, Anthropic, Google, and Meta have also reported incidents involving their models conducting autonomous hacks against other organizations.
The incidents prompted a UN Security Council meeting on AI security risks this week, where OpenAI CEO Sam Altman and Anthropic CEO Dario Amodei called for greater global cooperation on AI safety.
Twitter Post
Altman's post on the matter
There is an extensive and ongoing review related to our agents’ use of internet access during training and evaluation. We’ve been publishing summaries at the link below and will continue to.
— Sam Altman (@sama) September 25, 2026
We have not been as fast as we would have liked but we are trying to balance our desire… https://t.co/8zoMxas5Eq