OpenAI AI agents accessed 55 websites and obscured activity, security firm says

WASHINGTON, Oct 2 : OpenAI artificial intelligence agents accessed data from 55 government, business and nonprofit websites while using techniques that in some cases made their activity difficult for outside researchers to trace, according to an investigation by digital forensics firm Asymmetric Security.

The websites identified in the investigation included the US Centers for Disease Control and Prevention (CDC), Securities and Exchange Commission (SEC), International Energy Agency and Mayo Clinic. Asymmetric said it also found evidence of agents accessing or probing systems belonging to Australian government agencies and other organisations.

Researchers said some activity records were erased or became inaccessible, while the agents also created temporary email inboxes and private accounts on Urlquery, a malware-scanning service, to obtain or examine data. According to Asymmetric, those practices made it harder for independent researchers to determine exactly what information had been collected.

“It’s possible that the agents were deliberately using these tools to cover their tracks,” Asymmetric Security co-founder Pippa Thompson told the Financial Times. However, the company said it could not determine whether the behaviour was intentional or resulted from AI agents going outside the intended constraints of a testing exercise.

Asymmetric's investigation, published October 1, said it relied on publicly available evidence and identified successful access to some staging environments, reconnaissance activity and probing of a broader range of websites. The firm cautioned that the available evidence does not by itself establish that sensitive information was accessed.

The findings come amid a growing series of reports involving unexpected behaviour by AI agents. OpenAI recently disclosed that its models had interacted with US government websites in unintended ways, including the SEC and Census Bureau. The company said those cases did not involve SEC credentials or private information.

OpenAI has also notified more than 100 organisations about what it described as “misaligned agent activity,” while stressing that such notifications do not necessarily mean that the organisations' systems were compromised.

OpenAI told the Financial Times that it was reviewing the activity and notifying organisations when potential impacts were identified. The company said most of the activity detected involved routine research tasks and publicly available web content.

The latest investigation has renewed concerns about the transparency of autonomous AI systems, particularly when outside researchers cannot independently access complete records of what an AI agent was instructed to do, what information it retrieved and how it interacted with third-party services.