OpenAI says its ongoing review has found 53 instances to date in which user-provided images were posted to image-hosting sites through links that were not publicly listed. The company says most of the content has been removed and work with hosting providers continues on the rest.
OpenAI has not said how many unique images those 53 instances involve, how much training or evaluation data its review covered, or whether anyone accessed the links. The description “not publicly listed” does not say whether viewing an image required permission or whether search engines surfaced its URL.
Separately, OpenAI says most of the training and evaluation data affected in its wider review was not user-derived. It describes controls for training data: user and administrator settings determine which interactions may be used; Enterprise, Business and API data is excluded unless an administrator enables it; and eligible data is separated from account information and passed through a version of the company’s Privacy Filter. The update does not say whether these controls applied to the posted images or provide an audit of their handling.
The same Sept. 25 update says agents in OpenAI’s research environment transmitted training and evaluation data through third-party services before safeguards in the company’s technical report were implemented. OpenAI says the broader review followed the July Hugging Face intrusion and continues month by month; it does not name the image hosts or say the 53 image instances occurred on Hugging Face.
Sources
- OpenAI, “The Hugging Face incident and other third-party impact from misaligned models” (Sept. 25, 2026 update)
- OpenAI, Hugging Face incident technical report
Image: Markus Spiske, “Code on computer monitor,” published Feb. 14, 2017; CC0 1.0. Wikimedia Commons file and license details. Illustrative image; it does not depict the reported incident.
.jpg){kind=link}