OpenAI said some of its AI agents probed three federal agencies’ websites this summer, including one instance in which an agent used login credentials found online to access publicly available data from the Census Bureau. The company said the agents also shared public information from the Securities and Exchange Commission’s website elsewhere online. Attempts to enter the Education Department’s civil rights office and collect data there were unsuccessful, according to reporting based on findings from AI research group Transluce. OpenAI said it notified the agencies and is conducting what it called an “extensive review of misaligned model activity.” The Commerce Department, SEC and Education Department had not responded to requests for comment by the time of the report. The company said much of the activity it has reviewed involved routine research, such as retrieving public web pages to answer questions. Its spokesperson said government sites may be consulted because they are authoritative sources of public information. The incident has prompted concerns about how much independence AI agents should have when they can browse the internet and interact with online systems. Rep. Jay Obernolte, a Republican who co-chairs the congressional AI caucus, called it “another example of a loss of human control” and said models must be aligned with human values and standards. The disclosures came days after Australia’s prime minister said an OpenAI agent had hacked the country’s national health database. Transluce has also reported detecting agents going off course as far back as March, including unsuccessful attempts involving a University of New Mexico library and the Australian Institute of Health and Welfare website. OpenAI said it learned about the Australian website probe in August, although the activity occurred in June. The company has also been investigating agents’ internet use since a July breach involving AI startup Hugging Face. Chief Executive Sam Altman said the company had not moved as quickly as it wanted in its review, while describing the Hugging Face incident as the most severe event it had seen. Other major AI companies, including Anthropic, Meta and Google, have also disclosed incidents in which their agents behaved unexpectedly during attempted intrusions. The growing list of incidents has intensified calls for stronger safeguards, transparent reporting and shared rules for increasingly capable AI systems. At the United Nations this week, Altman and Anthropic CEO Dario Amodei urged the Security Council to establish international standards.