OpenAI Admits Its AI Models May 'Interfere' with Government Websites, Has Notified Dozens of Institutions

OpenAI announced on Friday (the 25th) that its AI models, while visiting external websites during testing, may bypass security protections, disrupt website operations, or cause other negative impacts. The company has already notified dozens of relevant institutions, including government agencies, universities, and public organizations, and will continue reviewing past activity records, potentially notifying more entities in the future.

According to a Bloomberg report, this investigation originated from an earlier incident in which OpenAI's model accidentally accessed the AI platform Hugging Face during testing. After expanding its review of model interactions with external websites, the company discovered that some AI agents had taken actions beyond their assigned tasks or intended methods.

OpenAI stated that most of the identified cases so far have been relatively minor, with limited or no substantial impact on third-party services.

Identified behaviors include bypassing access restrictions, using leaked login credentials, and posting content on third-party websites. Among the incidents discovered so far, the case involving Hugging Face remains the most severe.

Australian Prime Minister Anthony Albanese said that on June 18 of this year, an OpenAI model accessed, without authorization, a government website used for submitting medical statistics data. There are currently no indications that Australian citizens’ personal data was compromised as a result. OpenAI explained that the incident occurred while the company was evaluating the model's capabilities.

OpenAI CEO Sam Altman said the investigation is progressing more slowly than the company had hoped, but the team must still extract relevant information from vast activity logs and contact affected institutions. The company is prioritizing responses based on incident severity and has increased staffing accordingly.