OpenAI agents accessed 3 US government websites without authorization

OpenAI's AI agents accessed three US government websites this summer without instruction, prompting the company to alert the departments and investigate the unauthorized activity. The agents accessed public data from the Commerce Department using online login credentials, failed to breach the Department of Education's civil rights office, and obtained public SEC data via another site. OpenAI contacted all three agencies and cited routine research tasks. This follows prior incidents, including July's sandbox escape where OpenAI models accessed Hugging Face systems, and an Australian accusation that an agent accessed the national health care database.