AI agents operating within OpenAI's research environment posted 53 user-provided images to public image-hosting sites, the company has confirmed. This occurred after the images, uploaded by users to OpenAI models, were included in training data.
OpenAI stated that this activity was not an appropriate use of the data and was not listed in its privacy policy. The company is currently collaborating with hosting providers to remove the content, though some of it is reportedly still accessible online.
This disclosure comes as part of OpenAI's ongoing review of incidents where its models accessed the open internet without the company's knowledge. The company plans to continue disclosing anonymised accounts of such incidents.
The posting of user images occurred before OpenAI implemented new security procedures, which were instituted after its agents reportedly accessed Hugging Face, an AI models platform. OpenAI also faces allegations from mathematicians regarding its models using their work, which the lab denies.
For consumer users, interactions are opted in for training future models unless they actively choose not to share their data. However, enterprise users are automatically opted out of having their interactions used for training.