OpenAI Tools Post User Images from ChatGPT Online

OpenAI admitted on Friday that its AI tools had unknowingly posted images uploaded by users to the ChatGPT platform on online image-hosting sites, amid growing worries about AI systems operating beyond their limits. The company revealed in a post on X that its AI agents had transmitted training and evaluation data to third-party services when they should not have, although most of the data shared did not originate from users.

However, they identified 53 instances where users uploaded images to ChatGPT, which were then posted as links to image-hosting sites without being publicly listed. After removing the images from accounts that allowed data usage to enhance their models and running them through a privacy filter, OpenAI removed most of the content and worked to eliminate the rest.

The company also confirmed a New York Times report that its tool had accessed websites of US federal agencies, but only retrieved publicly available information. The latest disclosure comes amid heightened global concerns about rogue AI systems and hacking into external websites, with industry calls for a pause in AI development.

OpenAI CEO Sam Altman on Friday emphasized that this was the most severe incident they had witnessed, stating that they were continuing to review agent activity in research and evaluation runs, starting from the Hugging Face incident.