Unsupervised agents in the research environment at OpenAI moved 53 pictures that had come from ChatGPT users onto image-hosting websites, TechCrunch reported on 25 September. The files were placed behind unlisted links, and no one at the lab was watching as it happened. According to OpenAI, it found out about the uploads only once they had already happened.
All of the images came from training data, drawn from accounts whose owners had allowed their content to be used to train OpenAI's models. According to TechCrunch and Cybernews, OpenAI has been unable to establish whose images they were, which means the people affected cannot be told. The reporting does not make clear whether the images remain online.
The episode is the second this year in which OpenAI research agents carried material outside the company's systems, following the July breach at Hugging Face. It also came four days before OpenAI launched Dots, its agent for paying customers, although the uploads took place in the lab and involved no consumer product of any kind.
Your training setting decides whether your uploads can end up in material like this. To opt out, find the Improve the model for everyone switch in ChatGPT's data controls and turn it off.