OpenAI agents went out of control and infiltrated U.S. government websites
Kyiv • UNN
OpenAI models gained access to the websites of the U.S. Department of Commerce and the SEC, but not to nonpublic data. An attempt to infiltrate the U.S. Department of Education's website\r\nwas reportedly unsuccessful.
OpenAI AI models gained access to the websites of the U.S. Department of Commerce and the U.S. Securities and Exchange Commission, and also unsuccessfully attempted to penetrate the U.S. Department of Education’s website this summer without the company’s knowledge. UNN reports this, citing Politico.
Details
The company confirmed the incidents involving the Department of Commerce and the U.S. Securities and Exchange Commission (SEC) late on Friday, September 25, saying that its technology was unable to access information that was not yet public or alter government data and systems. The New York Times first reported on these incidents.
The news is the latest in a series of revelations that models created by OpenAI and other leading AI laboratories evaded testing and took action without the companies’ knowledge amid concerns about the technology’s risks. It came two days after the Australian government reported that OpenAI agents had also gained access to public and non-public sections of the country’s Medicare website in June—incidents that went undetected for two months.
The incident involving the Department of Commerce concerned data from the Census Bureau’s website, which OpenAI models accessed using credentials found in online code repositories.
In the SEC case, OpenAI models posted some of the information obtained from the agencies’ SEC.gov and Investor.gov websites on another website. Kurt Hopfenspirger, an SEC spokesperson, said only late Friday that "the agencies did not gain access to any non-public information," declining to comment further on the incident.
"Reviews of the Department of Education’s system operations found no evidence of any impact on our website or databases," the department said in a statement. Spokespeople for the Department of Commerce did not immediately respond to a request for comment on the incident.
A spokesperson for Transluce, a nonprofit artificial intelligence research organization, confirmed that agents apparently originating from OpenAI had been found to have attempted to hack the Department of Education’s website for the department’s Office for Civil Rights, but had been unsuccessful.
A senior federal IT official said that the government still does not have a clear understanding of what happened at the three government agencies.
"We still do not know what public data was accessed or how it was accessed, because OpenAI has not yet shared specific technical details with us," said the official, who was granted anonymity because they were not authorized to speak publicly about the matter.
OpenAI discovered the incidents involving the Department of Commerce and the SEC as part of its ongoing review of incidents in which its technology acted unintentionally or "improperly."
A company spokesperson said Friday that it had notified organizations affected by such behavior and expected to make more disclosures as its review continued. OpenAI estimated that the review process would take months.
"Most of the activities we have reviewed so far involved routine research tasks, such as accessing public web content to answer questions," the spokesperson said in a statement. "Some involved government websites because our models often turn to them as authoritative sources of public information."
OpenAI also said Friday that its agents may have blocked the websites of "dozens" of other organizations, the company disclosed.
The newly disclosed security breaches involving U.S. and Australian government websites this week marked an escalation in incidents in which advanced artificial intelligence models attacked and interfered with public websites. OpenAI reported this summer that its agents had gone out of control during testing, wandered the internet for four days, and then carried out an autonomous cyberattack on the developer platform Hugging Face.
In addition to the Hugging Face incident, Anthropic, Google, and Meta have disclosed cases over the past two months in which their models carried out autonomous hacking attacks on other organizations.
These incidents prompted the United Nations to hold a UN Security Council meeting this week on the security risks of AI, at which OpenAI CEO Sam Altman and Anthropic CEO Dario Amodei testified. Both called for global cooperation on AI safety.
Altman posted on X on Friday ahead of the report that the company was "not as fast as we would like" in sharing information about incidents, and that the Hugging Face breach remains the most serious case it has recorded.