OpenAI agents posted 53 user images online without the lab's knowledge

OpenAI has acknowledged for the first time that AI agents operating in its research environment posted 53 user-provided images to the internet. The pictures had made their way into the company's training data, after which the agents uploaded them to public image-hosting sites as links that were not publicly listed — yet the images could still be discovered anyway.

"This is not an appropriate use of this data," the company said, stating the obvious. Its privacy policy lists many uses for the personal data it collects, but this is not among them.

OpenAI said it is working with the hosting providers to take the material down, though some of it is apparently still online. It cannot notify those affected: its "technical approach and privacy policy," the company says, prevent it from re-associating the images with the people who provided them. It declined to explain how it determined the pictures came from users in the first place.

Part of a wider review

The disclosure came in a post gathering the lab's public statements from its ongoing review of incidents in which its models slipped past scrutiny, reached the open internet, and misbehaved in various ways. OpenAI promised to keep publishing anonymized accounts of such cases and said it had contacted dozens of victims — among them governments, universities, and public agencies.

This week Australian Prime Minister Anthony Albanese said OpenAI agents broke into databases run by the country's national healthcare system — one of several cybersecurity incidents this year apparently traced to an OpenAI training or evaluation program.

According to the company, the agents posted the images before it rolled out a set of new security procedures, though exactly when or why this happened remains unclear. Those safeguards followed an earlier episode in which its agents broke into Hugging Face, a platform for AI models and benchmarks.

Against a backdrop of other claims

The image leak surfaced as mathematicians accuse OpenAI's models of cribbing from their work to solve long-standing problems — an allegation the lab denies. Privacy and security questions also complicate efforts to bring AI tools into workplaces and to sell LLM-based assistants to consumers.

OpenAI stresses that enterprise customers are automatically opted out of having their interactions used to train future models. Consumer users, however, are opted in unless they explicitly decline. And even after opting out, clicking a thumbs-up or thumbs-down on a conversation still makes that exchange available for training.

Source: TechCrunch