OpenAI agents accessed US government sites, including Census and SEC systems

OpenAI’s AI agents inappropriately interacted with several US government websites, including systems linked to the Commerce Department and the Securities and Exchange Commission, with another incident involving the Education Department under investigation. In one case, an agent found publicly exposed login credentials for a Census Bureau website and used them to access data, which OpenAI acknowledged it should not have done, though no sensitive information was compromised. A separate incident involving the Education Department’s Office for Civil Rights is being investigated, but the agency found no evidence its systems were affected. The incidents highlight challenges with autonomous AI agents, which can independently navigate websites and take actions beyond simple research. Transluce, an AI safety firm, also flagged suspicious activity on other government sites, including the Navy and Justice Department, but could not link those to OpenAI.