OpenAI has confirmed that its artificial intelligence models accessed publicly available information from several US government websites, including the Census Bureau and the Securities and Exchange Commission, as part of a widening review of how its systems behave during training and evaluation.
According to a Bloomberg report citing people familiar with the matter, the company's agentic AI systems interacted with the SEC.gov, Investor.gov and Census.gov websites. OpenAI separately confirmed to Bloomberg that its models accessed public information from these sites during the training and evaluation of its technology.
On Friday, the company said it had informed "dozens" of organisations, including governments and universities, whose websites may have been affected by visits from its models during testing. In an X post on September 25, OpenAI said: "After the Hugging Face incident, we committed to conducting a much broader review of actions taken by our models during training and evaluation and to being transparent about our findings. This is an extensive review that is ongoing."
Review follows earlier breaches
The review traces back to an incident several months ago in which one of OpenAI's AI systems accidentally breached Hugging Face. In an extensive blog published on Friday, the company said it had contacted organisations about cases where its software may have bypassed security controls, disrupted a website's availability, or otherwise negatively affected external services.
The disclosure also follows OpenAI's recent admission that one of its models had accessed an Australian government website without authorisation earlier this year. Australian Prime Minister Anthony Albanese confirmed this week that OpenAI's technology had gained unauthorised entry into a government website used for reporting healthcare data in an incident on June 18, adding that it did not appear to compromise Australian citizens' personal information. OpenAI said the incident took place while it was testing its models.
Altman says review could take months
OpenAI chief executive Sam Altman said in a separate X post that the company had not moved as quickly as it would have liked. "There is an extensive and ongoing review related to our agents' use of internet access during training and evaluation," he wrote, adding that the company was "trying to balance our desire for transparency with gaining a clear understanding from petabytes of agent activity logs, and working with impacted organizations." He said cases were being prioritised by severity and that Hugging Face remained the most severe event seen so far.
Company spokesperson Liz Bourgeois told Bloomberg: "We're conducting an extensive review of misaligned model activity and notifying organisations when we identify potential impacts to their systems. We expect to make additional notifications as that work continues. Most of the activity we've reviewed so far involved routine research tasks, such as accessing public web content to answer questions."
Cybersecurity experts cited in the report warn the findings highlight a growing concern: AI systems can act without their creators' real-time awareness, and the models have shown the ability to identify previously unknown software vulnerabilities and exploit multiple weaknesses at once. Traditional cybersecurity tools such as firewalls and email filters, designed to detect known threats, are poorly suited to spotting such behaviour. The report added that similar incidents involving AI models from Anthropic, Google DeepMind and Meta have raised broader concerns among major technology firms.
Sources:
OpenAI Confirms Its AI Models Accessed US Government Websites, Including Census And SEC Data — ETV Bharat (via Bloomberg)
Last edited: