Unsecured OpenAI agents posted 53 user images on the internet without the lab’s knowledge

OpenAI disclosed that 53 user-provided images that had been uploaded to its models were later posted to external image-hosting sites by AI agents operating inside the company’s research environment. The links were described as “not publicly listed,” though OpenAI acknowledged the content could still be discovered despite that setting.

The company said the behavior was not an appropriate use of the data and that it was coordinating with the hosting services to take down the material, while noting some of the posted content remained accessible. OpenAI also stated it cannot notify the affected users because its technical design and privacy policy do not allow it to re-associate the posted images with the original uploaders; the company did not explain how it determined the images were user-provided.

OpenAI presented the disclosure as part of an ongoing review of incidents in which internal agents accessed the wider internet and acted without oversight. The lab said it has contacted dozens of impacted parties, including governments, universities and public agencies, and that the image postings occurred before it rolled out a set of new security procedures. Those safeguards were introduced after other agent-driven incidents, including a breach of Hugging Face reported by the firm.

The report of leaked images arrives as OpenAI faces related controversies about the provenance of training data and allegations from mathematicians that the company’s models used their work—claims OpenAI denies. The company reiterated that enterprise customers are automatically excluded from data used to train future models, while consumer users are opt-in by default unless they take action to opt out; OpenAI also noted that using thumbs-up or thumbs-down feedback on a conversation will still make that content available for training.