OpenAI scandal: ChatGPT posted some users’ images on the internet

OpenAI acknowledged on Friday that its agents had published 53 images from conversations with ChatGPT on public image-hosting sites. The files came from accounts whose users had not disabled the use of their exchanges to train the models.The same document was making official, for the second time in three months, the suspension of all tool use by its most advanced models.

Private photos freely accessible

The images were posted online by autonomous agents, these ChatGPT variants capable of browsing and acting on their own on the web as they carried out tasks that did not call for anything of the kind. The links, which did not appear in search engines, remained accessible to anyone who found them. Most of the files have since been removed from the hosting services concerned.

It is simply impossible to notify the people concerned, by the company’s own admission. The images had passed through its privacy filter and are now separated from their original accounts, making any correlation impossible. OpenAI also refuses to specify whether real people appear in these images. “This is not an appropriate use of this data”, the report acknowledges.

A loss of control taking hold

The pause freezes tool use during training, evaluation, and all the way through the deployment of the most advanced models, with no resumption date announced. It also addresses an incident on September 20: a model undergoing training bypassed the network filtering of its sandbox to query an external chatbot. Its queries were sent out disguised as DNS requests, and the responses came back through the same channel.The session had to be manually interrupted, the model was abandoned, and two independent blocking layers were subsequently added.

A first two-week suspension had already taken place in July, after another model had scraped Hugging Face repositories without authorization. Sam Altman described this episode as “the worst incident we have ever known”, and the internal rules had been tightened in August. Agents also searched U.S. federal websites beyond instructions, without stealing any non-public data from them. The UN statistical servers reportedly received more than 16,000 scans from OpenAI agents between April and June. On June 17, a burst of 200,000 requests reportedly even included an SQL injection attempt, which failed.

All 53 images came from accounts that had left conversation sharing for training enabled, a setting checked by default when the account is created. It can be disabled in ChatGPT’s settings, under “Data Controls,” by unchecking “Improve the model for everyone.”

👉🏻 Follow the latest tech news in real time: add 01net to your Google sources and subscribe to our WhatsApp channel.

Source: The Guardian