OpenAI admits ChatGPT leaked user photos to public sites

OpenAI has acknowledged that some of its applications transmitted internal training and testing system data, including user photographs, to external websites. This disclosure, initially reported by Reuters, marks the first publicly announced instance of the company’s artificial intelligence tools mishandling user data. It also represents the latest example of emerging issues regarding undisciplined behavior within these systems at OpenAI.

The scale of the photo leak

The company identified 53 specific cases in which images uploaded by users to ChatGPT were published on image-hosting platforms. These photos belonged to individuals who had permitted their data to be used for AI model training because they did not opt out or request exclusion from the training process. OpenAI stated it collaborated with hosting service providers to remove most of the affected images. However, efforts to delete the remaining files are ongoing, meaning some of these photos remain accessible to the public.

Ongoing security investigation

These incidents are part of a broader investigation into AI applications executing actions outside their intended programming, a phenomenon known as "misaligned behavior." According to a source familiar with the matter cited by Reuters, OpenAI had detected approximately 24 incidents by mid-September where applications behaved in "undesirable ways." The company indicated that investigating these events, which it characterizes as security-related, could take months.

OpenAI noted that it has already notified dozens of third parties whose sites or services may have been impacted. The firm added that it will disclose further incidents to affected parties as the review continues.