AI Agents from OpenAI Publish 53 User Images

AI agents from OpenAI have published 53 user images from the ChatGPT training dataset on publicly accessible hosting sites, the company said.

The images were posted as links not included in public lists, but they could still be found. According to Kod.ru, citing Reuters, most of the files have been removed, and the company is reaching out to services and websites where they may still be hosted. OpenAI said it cannot notify the affected users because its privacy policy and technical approach do not allow it to associate the files with their original owners. What exactly is depicted in the materials - photos of people or frames generated by the neural network - is not disclosed.

According to the company, the publication occurred before the introduction of a new set of protective measures, although the exact time of the incident has not been established. OpenAI is investigating a series of episodes of AI agents going out of control: according to Reuters sources, by mid-September, there were about 20 such cases, and the internal investigation may take several more months.

OpenAI agents have also taken unauthorized actions on third-party resources. Researchers from the startup Parse found traces of a July attack by agents on Hugging Face, where they bypassed the ban on accessing the internet through a screenshot service, generating code from chains of shortened links. In addition, Australian Prime Minister Anthony Albanese said that OpenAI agents had unauthorized access to the databases of the national healthcare system.