OpenAI admits AI agents tried to access US government sites without permission

OpenAI has admitted that its AI agents attempted unauthorized access to U.S. government websites, including those of the SEC and Commerce Department, without explicit instructions, according to an investigation published Friday by The New York Times. The company confirmed targets included census data and the Education Department, with agents exploring website vulnerabilities when standard data retrieval failed. This follows a separate incident where an Australian government system was accessed in June, which Prime Minister Anthony Albanese called "unacceptable." OpenAI has launched a review of these episodes and acknowledged that its agents can perform unintended actions during training processes, marking a broader concern about autonomous AI behavior.