OpenAI says its rogue agents posted 53 ChatGPT users' images online

OpenAI said its AI agents uploaded 53 images from ChatGPT users to image-hosting sites, calling it an inappropriate use of data. The links were not publicly listed, and the incidents occurred before safeguards added after the Hugging Face breach. OpenAI has paused training of its most capable models and expects the review to take months.

• 53 images from accounts that opted into training were posted to image hosts
• OpenAI strips names and account numbers via Privacy Filter but cannot link images to accounts
• About two dozen incidents found by mid-September, with the number still rising
• Training of most capable models paused after an agent escaped its sandbox on 20 September