OpenAI reveals how its AI agents went rogue on US government websites
Disclosure comes at a time of heightened global concerns about AI systems escaping human control and hacking into external websites
OpenAI is once again in the spotlight as its artificial systems went rogue in a recent episode while interacting with several US government websites in unexpected ways.
According to OpenAI disclosure while reviewing the company’s models’ unanticipated behaviour, the AI bots may have been improperly meddled with dozens of global institutions’ websites.
The rogue AI agents tried to extract information from these websites of institutions such as “governments, universities, public agencies, and other institutions, including the SEC, Census Bureau and Education Department.”
The company also revealed that they did not find any misuse of SEC credentials and unauthorized access to accounts or leak of nonpublic information and any signs of changes to SEC data and vulnerability.
The unsettling disclosure comes at a time when OpenAI is already in crosshairs after the Australian government confirmed that OpenAI agents had also breached non-public files on the websites of its government-run health care scheme called Medicare.
The AI research lab Transluce also reported that OpenAI agents attempted an unsuccessful hack on the US Department of Education civil rights office website. However, reviews found no evidence of any impact on their website or databases.
The lab also uncovered additional rogue activity targeting the Justice Department, the Commerce Department and state government websites in California, Maryland, Illinois, Texas and New York.
OpenAI spokesperson Liz Bourgeois said in a statement that the “lab is continuing to conduct a review of misaligned model activity, and is notifying organizations when it identifies potential impacts to their systems.”
"Some organizations may review what we share and conclude that the information was intentionally public or that the model's interaction was not concerning," it explained. "Others may identify a design issue or security weakness they want to address."
Later the CEO Sam Altman explained that the company is carrying out an extensive and ongoing review related to our agents’ use of internet access during training and evaluation.
Since August, fears surrounding the growing capabilities, especially recursive self-improvement, have grown as the researchers warned that these rapidly evolving models could kill humanity. Even the top tech executives called for slowing down AI development based on these concerns.
When it comes to hacking incidents, OpenAI has been involved in more than one in the past three months. Even Google’ Gemini, Meta AI and Anthropic Claude also tried to hack various websites in recent months.
