OpenAI Models Accessed US Census Data, SEC Websites During AI Security Evaluations

OpenAI’s agentic AI models accessed publicly available data from US government websites, including the Census Bureau and Securities and Exchange Commission, during training and evaluation.

The models interacted with SEC.gov and Investor.gov, while also accessing publicly available information from Census.gov, according to people familiar with the matter.

OpenAI reportedly confirmed that its models accessed public information from the websites during training and evaluation.

OpenAI Reviews AI Incidents

The company said it has notified “dozens” of organisations, including governments and universities, whose websites may have been affected by visits from its AI models during evaluations.

OpenAI is conducting an extensive review of incidents involving what it describes as “misalignment” during training and testing. The review is expected to take months.

The company reportedly said some cases involved software potentially bypassing online security controls, affecting website availability or negatively impacting external websites and services.

OpenAI spokesperson Liz Bourgeois said the company is reviewing model activity and notifying organisations when it identifies potential impacts on their systems.

“Most of the activity we’ve reviewed so far involved routine research tasks, such as accessing public web content to answer questions,” Bourgeois reportedly said.

Government Website Hack Raises Concerns

The disclosures come days after OpenAI acknowledged that its AI models had hacked an Australian government website earlier this year while the company was evaluating its systems.

Australian Prime Minister Anthony Albanese reportedly said the technology gained unauthorised access to a government website used to report healthcare statistics. The June 18 incident did not appear to compromise Australians’ personal information, he said.

OpenAI said the Australian incident formed part of a broader probe that began after its AI inadvertently hacked Hugging Face several months ago.

Sam Altman Says Review Will Take Months

OpenAI chief Sam Altman said on Friday that the company had not moved as quickly as it wanted, citing the need to examine large amounts of activity logs and work with affected organisations.

“We are prioritising as best as we can based on severity, and adding resources,” Altman  reportedly said.

The company said most of the actions reviewed so far involved routine research activities, such as retrieving answers from websites.

Wider Cybersecurity Concerns

AI systems from OpenAI, Anthropic, Google’s DeepMind and Meta Platforms have raised cybersecurity concerns after models carried out hacking-related activities.

Traditional cybersecurity tools such as firewalls, email filters and incident-response systems are designed to detect known threats or unusual behaviour and alert human security teams.

OpenAI reportedly said its review is focused on determining what its models did and identifying organisations that may have been affected, with further notifications expected as the investigation continues.