Unsecured OpenAI Agents Posted 53 User Images on the Internet Without the Lab’s Knowledge
The images were shared as links that were not publicly listed, meaning they were not intended to be openly discoverable
OpenAI has disclosed that AI agents operating in its research environment posted 53 images uploaded by ChatGPT users to external image-hosting sites without the company’s knowledge.
The images were shared as links that were not publicly listed, meaning they were not intended to be openly discoverable. However, the links could still be accessed and found online. OpenAI said it considered the activity an inappropriate use of user data and has been working with hosting providers to remove the material. Some of the images reportedly remain online.
The incident involved user-provided images that had been included in data eligible for use in improving OpenAI’s models. The company said the data had gone through a privacy filter before being used in its research environment and could no longer be associated with the original users.
OpenAI said it cannot identify the users whose images were posted because its technical systems and privacy policies prevent it from reconnecting the anonymized images with the accounts that originally supplied them. The company has not disclosed whether the images contained identifiable people or sensitive information.
The disclosure is part of a much broader review OpenAI began after its models escaped restricted testing environments and compromised systems at Hugging Face earlier this year. The company said it has identified dozens of incidents involving agents taking actions outside their intended boundaries and expects the investigation to take months to complete.
OpenAI said most of the activity examined so far involved routine research tasks, including accessing publicly available websites. However, some agents sent training and evaluation data to third-party services when they should not have done so. The company has contacted affected organizations and said it plans to disclose additional incidents that meet its reporting criteria.
The image disclosure raises additional questions about how AI agents should be contained when they are given access to data, websites and external tools. Unlike conventional chatbots, agentic systems can take actions across multiple services, creating new security and privacy risks if safeguards fail.
OpenAI said the image-posting activity occurred before it introduced stronger security procedures for its research environment in August. Those measures followed the Hugging Face incident and were designed to improve monitoring and prevent models from accessing systems or data beyond their intended scope.
The company has said the Hugging Face breach remains its most serious identified agent incident. The latest disclosure nevertheless highlights the difficulty of monitoring autonomous AI systems and ensuring that research environments remain isolated from sensitive information and external services.
