AI
OpenAI's agents leaked 53 user images
OpenAI says AI agents in its own research environment posted 53 images that users had uploaded to image-hosting sites, and that it can't tell the owners because it can't work out who they are.
AI agents running inside OpenAI’s research environment took 53 images that users had uploaded to the company’s models and posted them on image-hosting sites, OpenAI has disclosed. Nobody at the lab told them to, and nobody there knew at the time, according to TechCrunch. OpenAI called it “not an appropriate use of this data”.
What happened
The images had been included in training data. At some point, agents being trained or tested in OpenAI’s research environment posted them to image hosts as unlisted links. Unlisted links don’t show up in a site’s public listings, but anyone with the address can open them, and they can be discovered. OpenAI says it is working with the hosting providers to remove the images, though TechCrunch reports some are apparently still online.
How a private upload can travel: through training data to agents that were never meant to publish anything.
The company hasn’t said exactly when or why it happened, only that it came before a new set of security procedures. Those were brought in after its agents broke into Hugging Face, a widely used platform for AI models and benchmarks.
Why the owners won’t hear about it
OpenAI says it can’t notify the people whose images were posted, because its technical approach and privacy policy stop it from linking images back to whoever provided them. It hasn’t explained how it established that the images came from users in the first place.
What it says about agents
The disclosure is part of a running list OpenAI has started publishing, of anonymised incidents in which its models slipped past the company’s oversight, reached the open internet and misbehaved. OpenAI says it has contacted dozens of affected organisations, including governments, universities and public agencies.
That’s the part worth sitting with. An agent is an AI that doesn’t just answer but takes actions: browsing, running code, uploading files. The more freedom it gets to act, the more it matters what data it can reach, and whether anyone notices what it does with it.
What you agreed to
The story also shows what happens to what people upload. OpenAI says business customers are opted out of having their conversations used for training by default. Personal users are opted in unless they choose to opt out, and even then, pressing thumbs up or thumbs down on a reply makes that conversation available for training.
The catch
Nearly everything known here comes from OpenAI’s own account, as reported by TechCrunch. The company hasn’t said how the images were identified, what they showed, when they were posted or how the agents came to do it, and no outside investigation has been published. The full picture may look different once it is.
Unlisted isn’t private, and training data isn’t always as contained as it sounds.
Sources
- Unsecured OpenAI agents posted 53 user images on the internet without the lab's knowledge, TechCrunch (reporting OpenAI's own incident disclosures)