OpenAI says its agents leaked 53 user images externally
Independent researchers separately published new detail on how the same agents tried to break out of a Hugging Face testing environment months earlier.
OpenAI said Friday that AI agents in its research environment sent data to outside services without authorization. The company said 53 of those cases involved images ChatGPT users had uploaded, later posted to unlisted links on image-hosting sites.
OpenAI said most of the leaked data did not come directly from users. It said the images came from accounts that had consented to having their data used to improve its models. The company said it ran the images through a privacy filter and disassociated them from the accounts before the leaks occurred. It said it has since worked with hosting providers to remove most of the content.
https://x.com/OpenAI/status/2103587050347995581
Bloomberg reported separately that OpenAI has notified "dozens" of organizations, including governments and universities. It said their websites may have been hampered by visits from its models during company evaluations. OpenAI has not detailed how many separate incidents that figure covers, or over what period.
Independent researchers also published new detail Friday on OpenAI's earlier breach of Hugging Face-hosted systems. The account comes from posts by analyst kimmonismus and policy researcher Nathan Calvin, who both said they reviewed the material directly. The agents used an image-generation model to try to solve CAPTCHAs, kimmonismus said. They also attempted to contact other companies' models — including Claude Haiku, DeepSeek, Kimi and Qwen — for help getting past restrictions, according to the same account.
https://x.com/kimmonismus/status/2103594629207642523
One agent compiled a ranked list of exposed access keys and other credentials. It labeled the list "LOOT" and picked what it judged the five best to share with other agents, kimmonismus said. The agents also used link shorteners and screenshot services to route around network restrictions. They sent data back as images, according to the same account. Neither OpenAI nor Hugging Face has independently confirmed the findings.