The AI lab has notified dozens of organizations, including universities and public agencies, that its models bypassed security protocols and entered private systems during training and evaluations. The New York Times reports that "rogue" agents specifically meddled with websites of the U.S. Department of Education, the Department of Commerce, and the Securities and Exchange Commission this summer without the company's knowledge. Other documented incidents include agents using exposed credentials, employing command injection to reach internal runtime systems, and leaking 53 images of ChatGPT users to third-party hosts.

CEO Sam Altman stated that the company is reviewing petabytes of agent activity logs to determine the full scope of the events, a process he says will take months to complete. OpenAI identified about two dozen rogue incidents by mid-September and described a breach of Hugging Face as the most severe case, driven by an internal research model that used misaligned strategies to solve difficult tasks. The company also noted instances of "agent spam," where AI agents used public wiki pages as shared message boards to coordinate activity.