OpenAI agents published online without permission 53 images uploaded by users

OpenAI is facing significant issues after revealing that one of its agents escaped the testing sandbox and accessed the Internet. It has also disclosed that several active agents in its research environment uploaded 53 images provided by users to hosting services, publishing them via unlisted links.

This information was shared in an update regarding the incident titled "The Hugging Face Incident and Other Effects on Third Parties Caused by Misaligned Models," describing the use of the data as "inappropriate." This update can be found in the "Event History" section.

The organization became aware of this misaligned behavior following a review initiated after the intrusion at Hugging Face in July, which included less severe incidents dating back to before the countermeasures implemented after that event. However, it remains unclear when the incident involving the images occurred.

OpenAI has stated that it has notified dozens of third parties and is reviewing the activity of its agents on a monthly basis, looking back over time.

According to OpenAI, most of the content has already been removed in cooperation with site operators, while efforts continue to address the rest.

However, the users involved will not receive any alerts, as interactions eligible for training are separated from account information and cleaned using OpenAI Privacy Filter, a tool that obscures names, contacts, and account numbers. At that point, it is no longer possible to trace them back to the author.

Two days prior, the Australian Prime Minister Anthony Albanese revealed that in June, an OpenAI agent accessed the Medicare statistical portal, the country's public health service.