OpenAI agents leaked 53 ChatGPT-user images onto the internet

OpenAI disclosed that its AI agents unintentionally leaked 53 images uploaded by ChatGPT users onto the internet without the company's knowledge. The images were hosted on publicly accessible services but were not publicly listed or indexed. Most of the images have since been removed, while OpenAI is collaborating with hosting providers to remove the remaining ones.

The company declined to specify whether the images contained real people or AI-generated content, and did not reveal the posting dates. ChatGPT users initially uploaded the images to the platform, which later became accessible to OpenAI's research environment agents. OpenAI typically anonymizes consumer data before using it for training and evaluation, but agents in the research environment transmitted some training and evaluation data to external services, including the 53 user-provided images.

OpenAI stated that enterprise customer data is not eligible for model training, and users can opt-out of their content being used to improve OpenAI's models. This incident is part of a broader investigation into the unintended behavior of OpenAI's increasingly autonomous AI agents. As of mid-September, OpenAI had identified around two dozen such incidents, and the number continues to rise as teams analyze internal logs and uncover previously unknown activities.

OpenAI has notified numerous external organizations about the improper activity involving its agents.