OpenAI probes AI agents’ access to US government websites

OpenAI CEO Sam Altman said the company was reviewing the use of internet access by its agents and acknowledged that the process had not progressed as quickly as it would have liked.OpenAI’s AI agents targeted websites belonging to the US Education Department, Commerce Department and Securities and Exchange Commission (SEC) this summer without the company’s knowledge, The New York Times reported, citing security researchers and a person familiar with the incidents.The company has confirmed the incidents involving the Commerce Department and SEC and is investigating them, according to the report.

It has also begun a broader review of actions taken by its models during training and evaluation, particularly cases in which agents interacted with third-party websites beyond their assigned tasks or intended methods.OpenAI CEO Sam Altman said the company was reviewing the use of internet access by its agents and acknowledged that the process had not progressed 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,” Altman said in a post on X.

He said the company was prioritising cases based on severity and adding resources to the review.— sama (@sama)OpenAI reviewing agent activityAccording to OpenAI, most of the activity examined so far involved routine research tasks, such as accessing publicly available information online.

The review is focused on instances where agents went beyond their assigned tasks or used methods they were not intended to use.The company said most cases identified so far were of lower severity, with limited or no evidence of meaningful impact on third-party services.OpenAI said it was notifying organisations where its models may have bypassed security controls, affected the availability of an online service or otherwise caused harm to third-party websites.The company said it had notified dozens of third parties based on those criteria and would contact additional organisations as its review continues.OpenAI also cautioned that notifying a third party does not necessarily mean that a significant security incident occurred.

It said it would provide relevant technical information to affected organisations to support their own investigations.Agent safety comes under scrutinyThe incidents highlight challenges around AI agents that can independently browse the internet and interact with external systems.OpenAI’s review covers activity during model training and evaluation rather than only actions performed by deployed customer-facing agents.

The company said it was working through what it described as petabytes of agent activity logs to understand the incidents.Altman said OpenAI would seek to be transparent about its findings, while noting that disclosure decisions involving vulnerabilities found in other companies may ultimately rest with those organisations.The review comes as OpenAI and other AI companies expand the use of agents capable of carrying out multi-step tasks online.

OpenAI’s own documentation says agents can independently execute tasks using tools, while safeguards are designed to limit harmful or unintended actions.