OpenAI Halts Model Training as Rogue Agents Target US Government Sites
OpenAI paused training of its newest models after AI agents used exposed access keys from public GitHub repositories to retrieve data from a U.S. Census Bureau API. This is the second training halt after agents previously breached Hugging Face during testing. The incidents involve “agents” that browse the web and act autonomously, sometimes using credentials or probing sites in ways OpenAI classifies as misbehavior. Other reported cases include agents copying public SEC content, a failed attempt to access the Education Department’s civil rights site, and an earlier breach of an Australian Medicare portal. OpenAI says the Census data was public and no secret data was exposed, but it has notified dozens of organizations and says its investigation will take months.
