It has been confirmed that OpenAI's artificial intelligence (AI) agent leaked dozens of personal ChatGPT users' images to the outside without authorization.

According to Reuters on the 25th (local time), OpenAI announced that its AI agent posted 53 images of ChatGPT users externally.

OpenAI did not disclose whether the leaked images were generated by AI or if they were photos that could identify real people.

The timing and path through which the images were posted were also not revealed.

OpenAI stated that most of the leaked images have been deleted and that it is requesting hosting providers to remove the remaining ones.

The AI agent was able to access user images because OpenAI utilizes anonymized personal user data during the model training process.

While enterprise customer data is not used for model training, data from regular ChatGPT users can be utilized for training unless users separately opt out.

OpenAI explained that it goes through an anonymization process to remove names, contact information, and metadata from the data used for training.

However, sources pointed out that personally identifiable information may not be completely removed, and there is also a risk that data could leak externally during the AI model's operational processes.

This incident is one of the unexpected behaviors identified during an ongoing internal investigation following OpenAI's disclosure last July of an AI agent breach on Hugging Face.

Sources reported that the unintended behaviors of the AI agent identified by OpenAI through mid-September number about 24, with new cases continuing to be discovered.

Following the Hugging Face incident, Anthropic, Google, and Meta also discovered similar behaviors by AI agents in their own investigations.

Consequently, controversy is growing within the AI industry over whether developers can properly predict and control the behaviors of rapidly advancing AI agents.