OpenAI’s artificial intelligence models accessed publicly available information from US government websites, including those operated by the Census Bureau and the Securities and Exchange Commission, people familiar with the matter said.
The company’s agentic AI systems interacted with SEC.gov and Investor.gov and accessed publicly available information from Census.gov, according to the people, who asked not to be identified because they were discussing private matters. OpenAI separately confirmed that its models had accessed publicly available information from the government websites during training and evaluation.
The disclosure came as OpenAI expanded an investigation into potentially harmful activity by its AI systems. The company said on Friday that it had notified “dozens” of organisations, including governments and universities, whose websites may have been affected by visits from its models during evaluations.
In an extensive blog post, OpenAI said it had informed organisations about incidents in which its software may have bypassed an online service’s security controls, hampered availability, or caused a “negatively impacted” website or service outside the company. The incidents were identified as OpenAI broadened a probe launched after one of its AI systems inadvertently hacked the software-development platform Hugging Face several months ago.
OpenAI said affected websites included those operated by governments, universities, public agencies and other organisations. Liz Bourgeois, an OpenAI spokesperson, said the company was carrying out “an extensive review of misaligned model activity” and notifying organisations whenever potential effects on their systems were identified. “Most of the activity we’ve reviewed so far involved routine research tasks, such as accessing public web content to answer questions,” she said.
The investigation follows OpenAI’s acknowledgement earlier in the week that its AI models had hacked an Australian government website while the company was evaluating its systems. Australian Prime Minister Anthony Albanese said the technology had gained unauthorised access to a government website used to report healthcare statistics. The incident occurred on June 18 and did not appear to compromise Australians’ personal information, Albanese said.
OpenAI said on Friday that most of the activity it had reviewed involved models performing “mundane research tasks”, such as retrieving answers from websites. The company expects its review to take months.
OpenAI chief executive Sam Altman said the company had not moved as quickly as it would have liked but was attempting to balance transparency with the difficulty of identifying relevant incidents across large volumes of activity logs and then working with organisations that may have been affected. “We are prioritizing as best as we can based on severity, and adding resources,” Altman said in a post on X.
The disclosures come amid broader cybersecurity concerns surrounding AI systems developed by OpenAI, Anthropic, Google’s DeepMind and Meta Platforms. AI models have demonstrated an ability to identify previously unknown software vulnerabilities and, in some cases, exploit multiple flaws in combination to penetrate targeted organisations.
Traditional cybersecurity systems, including firewalls, email filters and incident-response tools, are generally designed to detect known malicious software or anomalous behaviour before alerting human security teams. AI-driven compromises can be more difficult to detect and potentially allow attackers to gain extensive access to systems while avoiding the attention of cybersecurity personnel.
