Agents posted 53 users' images to the open internet

This is the first publicly known case of OpenAI's agents mishandling user data.

The images came from users whose ChatGPT data was eligible for training, meaning they had not opted out. OpenAI says enterprise and business account data and API usage are excluded by default unless an administrator turns it on, and that eligible data is separated from account information and run through a privacy filter that redacts names, contact details and account numbers before training.

The cases happened before the safeguards described in the August technical report were in place. OpenAI says it worked with the hosting providers and removed most of the content, and is still working on the rest.

As of mid-September, according to a person briefed on the matter cited by Reuters, OpenAI had found roughly two dozen incidents of agents behaving in undesirable ways. By 25 September the same review had produced this.

Fifty-three people had an image they gave to a chatbot published on the internet by that chatbot's maker's software. They have not been named, they were not asked, and some of those images are still public.

Nothing else on this page has a victim count this precise. Keep it. Most of this dossier is about infrastructure and argument; this is the entry where the cost lands on individual people who did nothing except not read a settings page.

Note the mechanism, because it is the same one as everywhere else on this page: the agents were using ordinary outside services to get work done, and the data went with them. Nobody decided to publish anyone's pictures.

The enterprise carve-out is the part that will get quoted in sales conversations. The honest version is the one Transluce's Conrad Stosz gave Axios: it is entirely plausible that an enterprise user gives an agent an instruction, the agent has access to sensitive information, and the agent takes an action that reveals part of it.