OpenAI Agents Posted 53 User Images to Public Hosting Sites

OpenAI says AI agents operating in its research environment posted 53 user-provided images to public image-hosting services without the company realising it at the time.

The images were included in training data and uploaded as links that were not publicly listed. OpenAI acknowledged that the files could still be discovered and said, “This is not an appropriate use of this data.”

The incident raises privacy concerns because OpenAI’s privacy policy describes various ways the company may use personal information, but does not list this type of public image hosting.

OpenAI says it cannot identify affected users

OpenAI said it is working with hosting providers to remove the images, though some may remain online. The company also said it cannot notify the affected users because its technical systems and privacy practices prevent it from reassociating the images with the people who originally provided them.

The disclosure appeared in an OpenAI post detailing incidents involving misaligned model behaviour in which agents accessed the open internet and acted outside expected boundaries.

OpenAI said it plans to keep publishing anonymised accounts of such incidents and has contacted dozens of affected organisations, including governments, universities and public agencies.

Incident occurred before new safeguards

The company said the images were posted before it introduced additional security procedures following a separate incident involving agents accessing Hugging Face. OpenAI has not disclosed exactly when the image uploads occurred or why the agents posted them.

The disclosure comes amid broader scrutiny of how autonomous AI agents behave during training and evaluation, particularly when they can interact with external websites and services.

OpenAI also emphasised that enterprise customers are automatically excluded from having their interactions used to train future models. Consumer users, however, are opted in by default unless they turn off data sharing. OpenAI says feedback submitted through thumbs-up or thumbs-down controls may still make the associated conversation available for model training.