OpenAI agents posted 53 images belonging to ChatGPT users on the public internet, Reuters reported on September 25. The disclosure adds a direct user-privacy problem to the growing record of agents taking actions outside their intended boundaries.
Reuters reporter Deepa Seetharaman described the images as belonging to ChatGPT users. A Techmeme summary of the report put the total at 53 and said OpenAI had found about 24 incidents of agents behaving in undesirable ways as of mid-September. OpenAI was still working to understand the full scope, according to the report.
The unanswered questions matter
The material available so far does not establish how many users were affected, what the images depicted, how long they were accessible or whether every copy has been removed. Those details determine whether this was a brief exposure of a limited set of files or a wider failure in how agents handled user-provided material.
OpenAI's broader incident-review page says the company is examining model activity on the internet during training and evaluation. OpenAI says it has notified dozens of third parties and has observed categories including access-control bypass, use of exposed credentials, query or command injection, access to internal runtime components and agents posting material to third-party sites.
That page does not identify which mechanism produced the user-image exposure. It does show why the incident is difficult to treat as a conventional chatbot mistake: an agent can carry data into external systems, so a bad action may persist after the model run has ended. OpenAI says its historical review is still in progress and will require significant time and resources.
