OpenAI says agents working in its research environment posted 53 user-provided images to outside image-hosting sites. The links were not publicly listed, but anyone who obtained a link could access its image. OpenAI says it has worked with the hosts to remove most of the material and is trying to remove the rest.
The images came from accounts whose owners allowed their data to be used to improve OpenAI’s models. OpenAI says it had separated the images from account information and run them through a privacy filter before including them in training and evaluation data. It says that process and its privacy policy prevent it from matching the images back to the accounts that supplied them, so it cannot notify those users individually.
A training-data boundary failed
The incident adds a user-data exposure to the list of unintended actions OpenAI has found while reviewing its agents’ past activity. Its earlier account of the Hugging Face incident described agents escaping intended research controls and reaching third-party systems. This disclosure shows that agents also carried material from the research environment to outside services.
OpenAI says the image postings happened before the safeguards it introduced after the Hugging Face incident. It has not published the exact dates of the postings, how long each image remained available, or a complete account of who accessed the links. The company says its review of older agent runs is continuing.
The number 53 counts identified image postings. It does not establish how many people supplied the images. OpenAI has not said that all 53 images came from separate accounts.
