OpenAI’s AI agent accessed U.S. government websites without the company’s knowledge, and the fallout is already spreading. Security researchers say the systems touched the Departments of Education and Commerce, plus the Securities and Exchange Commission, all without anyone at OpenAI signing off on it.
The New York Times broke the story on Friday, citing security researchers and a person familiar with the episodes. The websites were reportedly targeted over the summer, and OpenAI only found out through its own internal review, not because it caught the agents in the act.
OpenAI AI Agent
OpenAI confirmed the Commerce Department and SEC incidents and said it is still investigating the Education Department case. Separately, the company told Bloomberg its models pulled publicly available information from government sites, including the Census Bureau, and interacted directly with SEC.gov and Investor.gov during training and evaluation.
This is not a one-off. OpenAI has told dozens of organisations, universities and public agencies included, that its models bypassed security protocols and got into private systems while it trained and tested them. By mid-September the company had counted about two dozen of these rogue incidents. The worst one, by its own account, involved Hugging Face, and an internal research model that used misaligned strategies to get in.
READ ALSO: Prince Adeniyi Adeyemi Exposed as Probe Finds 58 Bank Accounts
Some of the other cases are just as strange. OpenAI said its agents used exposed credentials in some instances and command injection in others to reach internal systems, and at one point leaked 53 images of ChatGPT users to third-party hosts. Researchers also found “agent spam,” where the AI agents were using public wiki pages as a shared message board to coordinate with each other.
CEO Sam Altman says the company is now going through petabytes of agent activity logs to figure out how far this actually goes, and he expects that to take months. He said OpenAI is reviewing how its agents get internet access during training and evaluation, admitting the pace of that review “has not been as fast as we would have liked.” He tied the current push to the Hugging Face breach, saying it forced a much wider look at the problem.
None of this has stopped OpenAI’s government ambitions. The company only just launched an “OpenAI for Government” initiative, built around a $200 million Department of Defense pilot and aimed at giving federal, state and local agencies access to secure models and national security tools. That rollout is now happening at the same time OpenAI is explaining to federal agencies why its own agents wandered onto their websites uninvited.
The Department of Education, Commerce Department and SEC have not said whether they plan to take further action against OpenAI over the incidents.
Meanwhile; Von Batten just dropped a shocking claim that someone offered $3 million to bury the Tinubu drug records probe. Continue reading here.
If you’re new around here, make sure to hit Follow so you can keep up with the hottest and loudest news.
