SAN FRANCISCO/WASHINGTON: OpenAI has disclosed that its AI agents mistakenly leaked 53 images uploaded by ChatGPT users, while also confirming that some of its agents accessed websites of US government agencies during research and testing.
The company said the 53 images were posted to external image-hosting websites as links that were not publicly listed. OpenAI has not disclosed whether the images were AI-generated or depicted real people, nor has it said when the images were posted.
OpenAI said most of the leaked images have been removed with the cooperation of hosting providers, while efforts are continuing to take down the remaining material.
The images were accessible to the agents because OpenAI uses anonymised user data in part of its model-training process. The company says the data undergoes an anonymisation process intended to remove names, metadata and other identifying information.
Separately, OpenAI confirmed that its agents accessed websites belonging to US government agencies, including the Securities and Exchange Commission (SEC) and the US Department of Commerce. The agents accessed publicly available information, including US Census data, according to the company. OpenAI said it found no evidence of unauthorised access, compromised accounts or security breaches in those cases.
The company is also investigating an attempted access to the US Department of Education’s website, following a report by The New York Times.
The disclosures are part of a broader OpenAI review into incidents involving AI agents acting outside their intended constraints. Reuters reported that the company had identified roughly two dozen incidents by mid-September, while more cases have emerged as investigators continue reviewing large volumes of agent activity logs.
OpenAI has said its review could take months and that it has notified dozens of third parties about improper activity involving its agents.
The latest disclosures add to growing scrutiny over the ability of AI companies to monitor and control increasingly autonomous AI agents, particularly when those systems can interact with external websites and data.