OpenAI: Our tools mistakenly published images of ChatGPT users
OpenAI acknowledged that its artificial intelligence tools published 53 images sent by users via its ChatGPT platform on the internet, in a new leak that this time involves the company’s own user data.
The San Francisco-based AI giant also revealed that its tools accessed websites belonging to U.S. federal agencies, confirming information previously reported by The New York Times, but noted that it only collected publicly available information in those instances.
OpenAI clarified that the links to the 53 images were not publicly accessible and were mistakenly published on web hosting sites.
The images had been sent from accounts whose owners had consented to the use of their data to develop the company’s models. Before use, the data passed through a personal data privacy filter, and the company confirmed that the images can no longer be linked to the users who sent them.
OpenAI did not specify whether the images show identifiable individuals or sensitive data.
The images were published as a result of the behavior of “AI agents,” which are software designed based on AI models and capable of autonomously executing a sequence of tasks.
The company explained that agentic AI tools used within its research transferred training data to external services, most of which did not originate from users.
It noted that most of these transfers occurred before it strengthened security around its research in August.
OpenAI confirmed that it is currently conducting a comprehensive review of the activities of its previous agentic tools, a process that will take several months.
Company CEO Sam Altman admitted that “we were not as fast as we would have liked” in reviewing and disclosing incidents, justifying this by the enormous volume of data that needed to be analyzed.
An OpenAI spokesperson told AFP that “the majority of activities audited so far were within the scope of routine research tasks, such as accessing publicly available content on the internet to answer questions.”
This confirmed that the company’s models accessed public information on the websites of the U.S. Securities and Exchange Commission, the financial markets regulator, before publishing part of it on another internet page, while emphasizing that it found no evidence of access to non-public information.
