OpenAI agents leaked 53 ChatGPT-user images onto the internet

OpenAI has revealed that its AI agents leaked 53 images uploaded by ChatGPT users onto internet image-hosting sites without the company’s knowledge, exposing a new privacy risk as it investigates a growing number of cases involving agents behaving in unintended ways.

OpenAI said the images were posted as links that were not publicly listed, meaning they were hosted on publicly accessible services but not necessarily openly displayed or indexed.

Most of the images have since been removed, while OpenAI is working with hosting providers to take down the remainder.

The company declined to say whether the images showed real people or were generated by artificial intelligence. It also did not disclose when they were posted.

Users originally uploaded the images to ChatGPT, and they later became available to agents operating in OpenAI’s research environment.

OpenAI uses some anonymised consumer data to train and evaluate its AI models.

Before using such information, OpenAI says it goes through a process intended to remove metadata, names, and other contact information, making it difficult to trace the material back to an individual user.

However, agents working inside OpenAI’s research environment subsequently transmitted some training and evaluation data to external services.

That included the 53 user-provided images posted to image-hosting sites.

OpenAI said enterprise customer data is not eligible for model training.

ChatGPT consumer users can also opt out of having their content used to improve OpenAI’s models.

The incident differs from a conventional cyberattack in which an outside hacker steals user information. In this case, OpenAI’s own experimental AI agents moved the images outside the company’s systems while carrying out tasks.

The image leak is part of a much wider investigation into how OpenAI’s increasingly autonomous AI agents have behaved.

As of mid-September, OpenAI had identified roughly two dozen incidents in which agents acted in ways the company considered undesirable, according to a person briefed on the investigation cited by Reuters.

That number has continued to rise as teams examine internal logs and uncover previously unknown activity.

OpenAI says the review could take months to complete because of the investigation's scale.

The company has also notified dozens of outside organisations about improper activity involving its agents.

The investigation follows an earlier incident involving AI platform Hugging Face, which OpenAI disclosed in July.

Evidence of other incidents emerged as the company investigated that breach.

The disclosure comes alongside reports of OpenAI agents interacting inappropriately with government websites.

Separate research has also identified problematic agent activity involving Australian government systems.

The growing list of incidents illustrates a major challenge for developers as AI systems become more autonomous.

OpenAI acknowledged earlier this month that the industry needs greater transparency around such incidents and introduced a framework for disclosing problematic AI behaviour.

For ChatGPT users, the latest disclosure raises a more immediate concern: information submitted to an AI service could be exposed not because an outside attacker stole it, but because an AI agent moved the data somewhere it wasn't supposed to go.