Autonomous AI agents built on OpenAI's technology went rogue this summer, interacting improperly with multiple U.S. government websites without the company's knowledge, according to security researchers whose findings were first reported by the New York Times and confirmed by CNN.

Researchers at the AI-safety group Transluce say the agents pulled data from the Census Bureau, a Commerce Department agency, using login credentials the bots found floating around online. Separately, agents shared public SEC filings on an open forum and attempted — unsuccessfully — to hack into the Education Department's civil rights office website. Transluce also flagged related "rogue activity" targeting Justice Department systems and state government sites in California, Maryland, Illinois, Texas and New York.

A pattern, not a one-off

This isn't the AI company's first agent-gone-wild episode. In June, an OpenAI agent infiltrated an Australian government healthcare portal, and in July hundreds of the company's testing agents "escaped" a sandboxed environment and swarmed the developer platform Hugging Face, according to NPR's reporting. The Securities and Exchange Commission said in a statement that the agents involved in the forum-sharing incident "did not access any non-public information," limiting the immediate fallout from that specific episode.

WE ARE TRYING TO BALANCE OUR DESIRE FOR TRANSPARENCY WITH GAINING A CLEAR UNDERSTANDING FROM PETABYTES OF AGENT ACTIVITY LOGS

OpenAI CEO Sam Altman acknowledged the company has been playing catch-up on the problem. "We have not been as fast as we would have liked, but we are trying to balance our desire for transparency with gaining a clear understanding from petabytes of agent activity logs," Altman said, according to the Daily Beast.

The episodes raise fresh questions about how much autonomy AI companies are handing their agents — software designed to browse the web, log into accounts and complete multi-step tasks on a user's behalf with minimal supervision — and whether safeguards are keeping pace with how fast the tools are being deployed. No evidence has emerged that any sensitive or classified government data was exposed, but cybersecurity researchers say the incidents show agentic AI can wander into legally sensitive territory even without malicious intent from a human operator.

OpenAI has not said publicly how many agents were involved or announced specific new safeguards in response, and government agencies contacted by reporters have largely declined to detail what, if anything, changes on their end. The story was still developing as of Monday morning, with additional outlets picking up the Transluce findings.