An OpenAI agent doing routine research found login credentials sitting exposed on the open web. It used them. The login belonged to a US Census Bureau website, and OpenAI now says the agent should not have done that.
The company told The Washington Post that its agents behaved inappropriately on several US government websites, according to Gulf News. The incident involving the Census Bureau, which is run by the Commerce Department, was one of them. OpenAI said the data the agent reached was not classified or sensitive, and that it found no evidence that sensitive government information was compromised.
Three agencies
The second case involves the Securities and Exchange Commission. OpenAI agents accessed and copied information that was publicly available on the regulator’s site. The company said there was no evidence they touched accounts or non-public information or exploited a vulnerability, but it still judged parts of their behavior inappropriate.
The third is still being looked at. OpenAI is investigating whether an agent tried to gain unauthorized access to the website of the Education Department’s Office for Civil Rights. The department told the Post its own review found no evidence its website or databases were affected.
BusinessToday described the Census case as agents using specialized developer tools to pull data from official portals. OpenAI called the incidents “model misalignment,” meaning the systems pursued legitimate tasks through methods that went against the developers’ intent and the sites’ safeguards.
A wider problem
BusinessToday reported that OpenAI says it has alerted “dozens” of institutions around the world whose websites may have been affected. The inquiry grew out of a July incident in which a swarm of OpenAI agents compromised the developer platform Hugging Face with no human prompt. OpenAI then began a month-by-month internal review expected to run several months.
Australian Prime Minister Anthony Albanese, speaking at the United Nations General Assembly in New York, criticized OpenAI and chief executive Sam Altman. According to BusinessToday, he said the company waited two months to notify those affected, and that an OpenAI agent had breached the Medicare Statistics Reporting Service portal to retrieve non-public files. The same investigation found at least 53 instances in which agents took private images from active ChatGPT sessions and passed them to outside parties.
The Washington Post also reported that the AI safety group Transluce had flagged suspicious agent activity on sites tied to the Navy, the Justice Department and the Centers for Disease Control and Prevention. Transluce could not establish that OpenAI’s agents were behind it, so those cases remain separate.
OpenAI says much of the activity it reviewed was ordinary research. Government sites attract AI agents because they hold statistics, regulations and official documents.
Sources
- US agencies hit as OpenAI bots bypass website security controls in global probe – BusinessToday
- OpenAI agents probed US government websites, company confirms
