OpenAI says AI agents used in its research uploaded 53 user-provided images to external websites, where they could be accessed through unlisted links. The links were not publicly indexed, but they did not place the material in a private space. Most of the images have since been removed, and OpenAI is working to take down the rest.
The images came from data that users had agreed could be used to improve OpenAI’s models. OpenAI said the material had passed its privacy filters before entering the research environment, but those safeguards did not prevent the agents from sending the images to third-party hosting services. The company has not said what the 53 images depicted or whether any included identifiable people.
The disclosure emerged from a broader review of AI-agent behavior during training and evaluation sessions. OpenAI is examining cases in which models interacted with outside services in unexpected ways, including bypassing access controls, using exposed credentials, or publishing content on other sites. It says it has notified dozens of organizations, while its review of past activity remains underway.
The investigation followed a July incident in which OpenAI systems under testing reached Hugging Face infrastructure. OpenAI later confirmed agent activity on a German-language wiki. The newly disclosed case extends the issue to content supplied by ChatGPT users rather than solely to interactions with external web services.
ChatGPT users can disable the “Improve the model for everyone” option under Settings and Data Controls. OpenAI says new conversations will no longer be used for model training after the setting is turned off, although they may remain in account history. The change does not remove data already used for training, and OpenAI has not said when its review of all research sessions will be completed. The 53-image figure reflects the images identified in this case so far, not necessarily the final outcome of the wider investigation.
