OpenAI's models accessed US govt websites including US census and SEC data

The company said its models accessed publicly available data from Census and SEC websites as it investigates incidents involving AI systems bypassing security controls

By Rachel Metz, Jeff Stone and Jake Bleiberg

 

OpenAI’s artificial intelligence models accessed publicly available information from US government websites, including those of the Census Bureau and the Securities and Exchange Commission. 

The company’s agentic AI systems interacted with SEC.gov and Investor.gov, as well as publicly available data from Census.gov, people familiar with the situation said, asking not to be identified, discussing private matters. OpenAI separately confirmed that its models accessed publicly available information from the websites during training and evaluation of the technology.

 

Earlier Friday, the company said it had notified “dozens” of organisations, including governments and universities, whose websites may have been hampered by visits from its AI models during evaluations of the technology. 

 

 

OpenAI has said it’s carrying out an in-depth review of incidents of “misalignment” during training and testing, which will take months to complete. The company’s efforts to work out exactly what its AI agents have done underscore one of the biggest concerns among cybersecurity experts about these incidents: The systems have carried out these tasks largely without the companies that created them knowing in real time.

 

In an extensive blog post Friday, the company said it had notified a range of groups about cases in which its software may have bypassed an online service’s security controls or hampered its availability, or in situations where a misaligned AI model may have “negatively impacted” a website or service outside of OpenAI. The company discovered these incidents while expanding a probe it began after its AI inadvertently hacked Hugging Face several months ago.

 

OpenAI said the websites that were affected include ones run by governments, universities, public agencies, and other groups.

 

In a statement, OpenAI spokesperson Liz Bourgeois said, “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.”

 

Only days ago, OpenAI acknowledged that its AI models hacked an Australian government website earlier this year, marking one of the first known AI cyberattacks on a government database. The company said in a statement that the breach occurred while it was evaluating its models.

 

Australian Prime Minister Anthony Albanese said this week that OpenAI’s technology had gained unauthorised access to a government website used for reporting healthcare statistics. The hack, on June 18, didn’t appear to compromise Australians’ personal information, he said. 

 

OpenAI said a post on social network X on Friday that most of the actions it has reviewed involved AI models carrying out “mundane research tasks,” like getting answers to questions from websites. The company said it expects it to take months to finish its review.

 

In a post on X Friday, OpenAI chief Sam Altman said the company has not worked as fast as it would like, but that it is balancing transparency with the need to find information in huge amounts of data in activity logs and then work with the companies that were impacted.

 

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

 

Hacks by models from OpenAI, Anthropic PBC, Google’s DeepMind and Meta Platforms Inc. have resulted in widespread cybersecurity concerns for major companies. Cyber vendors typically provide products that monitor for known strains of malicious software, or detect and block anomalous behaviors. Traditional cyber software such as firewalls, email filters and incident response tools specialise in detecting those threats, and then alerting human staffers who isolate breached accounts or devices.

 

AI models have proven to be significantly more advanced, sometimes finding previously unknown software vulnerabilities and then using multiple flaws at a time to breach a targeted organisation. Such compromises are harder to stop, and could provide malicious attackers with deep access to infected systems while remaining hidden from cybersecurity staffers.