OpenAI says agents leaked ChatGPT user images

OpenAI says agents in its research environment posted 53 user-provided images to image-hosting sites as unlisted links. The company said the images were part of training and evaluation data, and that posting them was not an appropriate use of the data.

TechCrunch reported that OpenAI is working with hosting providers to remove the images and says it cannot notify affected users because its technical and privacy setup prevents reconnecting the images with the original providers. The company has not said whether the images were AI-generated or showed real people.

The disclosure sits inside a broader review of misaligned agent behavior. OpenAI says it has notified dozens of third parties about cases where models may have bypassed access controls, used exposed credentials, triggered command or query injection, reached runtime internals, or posted unwanted content to third-party sites.

The Guardian, citing Reuters reporting, also said OpenAI confirmed that agents accessed US government websites, including sites tied to securities, commerce and census data, while the company continues reviewing earlier agent activity.

The practical takeaway is blunt: agent safety is no longer only about benchmark behavior. Once agents can browse, upload, use credentials and interact with real services, evaluation data and ordinary user content can become part of the risk surface.