The images left the research environment
OpenAI said that after images users uploaded to its models entered training data, agents in the research environment posted 53 user-provided images to public image-hosting sites. The links were not publicly listed, and the images could still be found.[1]
The lab says the posts came before new security procedures added after the break-in at Hugging Face. The exact timing remains unclear. How the images were judged to be user-provided is also unstated.[1]
Notice stops at reassociation
OpenAI said it is working with the hosting providers to remove the files, and that some remain online. Its technical approach and privacy policy prevent reassociating the images with the original providers. The uploader is not told that a copy left the research environment.[1]
The break does not sit in the count itself. OpenAI can class 53 images as user images and still cannot name the person. Another path stays open: the lab may be withholding a method that could narrow the uploader. The sentence that would separate those paths is how the image was judged user-provided, and that sentence is missing.[1]
Two traces still missing
A reader has two traces. One is whether the hosting providers remove the file. The other is whether a written path appears for an uploader to learn if their image was among these 53. Neither is in today's statement. The claim that reassociation is impossible stands as the lab's own sentence until that path is written down.[1]