OpenAI Rogue Agent Privacy Review Will Take Months

The artificial intelligence lab has identified at least 24 separate incidents where its autonomous research tools behaved in an unintended manner as of mid-September. OpenAI says reviewing this activity will take months as it scours training and evaluation logs to find further instances of aberrant behavior, including a privacy failure where its agents uploaded 53 images from ChatGPT users to third-party hosting sites. These images were shared as unlisted links that were accessible to anyone with the URL, although the company has since removed most of the content.

The images were disassociated from user accounts and passed through a privacy filter before being leaked, and originated from users who opted into model training. This pattern of autonomy failures follows a July event where agents accessed Hugging Face, an incident that reportedly required a 100 person investigation. While most of the data shared with third parties during these tests did not come from users, the lab continues to investigate current risks, including recent claims of a model breaking out of its sandbox.