There’s no end in sight to the saga surrounding the “violations” by OpenAI’s AI agents. In the latest episode, agents operating in the company’s research environment took 53 images that real users had uploaded to ChatGPT and posted them on image-hosting sites on the open web, without OpenAI being aware at the time of what had happened.
This incident is particularly significant, as this time the agents did not limit themselves to accessing external systems or exceeding the boundaries of a test environment. The images came from user data that had been transferred, after an anonymization process, to the company’s research environment for model training.
ΣΧΕΤΙΚΑ ΑΡΘΡΑ
OpenAI has not clarified whether the 53 images depicted real people or were generated by artificial intelligence, nor exactly when they were posted. According to the company, most of them have now been removed, and it is in contact with the hosting providers to have the rest taken down as well.
The company maintains that before user data is used for training, metadata, names, and other contact information are removed to make it difficult to link the data to a specific individual. Corporate customers are exempt from having their data used for training, while regular ChatGPT users must opt out themselves if they do not want their conversations and data used for this purpose.
In the case of the 53 images, in fact, OpenAI states that it cannot notify the users from whom they originated, because its data processing methods and privacy policy do not allow it to re-link the images to their original users.
From agents’ “escapes” into user data
This new revelation significantly expands the scope of the problem OpenAI is facing. By mid-September, the company had identified approximately 20 instances of unwanted agent behavior, and the number continued to rise as researchers examined the logs of their activity. OpenAI estimates that a full review will take months and reports that it has already notified dozens of third-party organizations about problematic agent behavior.
At the same time, it was revealed that the company’s agents had interacted with websites of U.S. government agencies, including the U.S. Securities and Exchange Commission (SEC), the Census Bureau, and the Departments of Commerce and Education. In the cases of the SEC and the Census Bureau, the information they accessed was publicly available, while OpenAI is separately investigating a possible attempt to breach the Department of Education’s system.
The incidents that have come to light so far—including those on Hugging Face and in government systems— had primarily raised the question of whether agents can overcome the technical constraints imposed on them and act outside the environment in which they are tested. The leak of the 53 images now directly raises the issue of user privacy as well.
OpenAI’s internal investigation is also under scrutiny. News agency sources describe it as strictly limited by the company’s legal team. Approximately 100 people are reported to have been involved in some way in the investigation of the incident at Hugging Face, during which evidence of other cases began to emerge. OpenAI denies that its lawyers attempted to prevent a broader investigation.
Two months after the initial revelation that OpenAI’s agents had strayed beyond the boundaries of their testing environment, the company is therefore still trying to assess the scope of the problem. Except this time, among the data found outside those boundaries was material that ChatGPT’s users themselves had provided to the system.
Source: ERT
