According to Reuters reports on September 25th, local time, two people familiar with the matter said that OpenAI is still investigating the extent of the abnormal activities of its AI agents.
On September 25th, OpenAI stated that its AI model had recently leaked 53 images of ChatGPT users. The company refused to explain whether these images were generated by AI or contained photos of real people, and did not disclose the timing of the images’ release.
As of mid-September, a person familiar with the matter estimates that OpenAI has identified approximately 24 instances where its intelligent agents have engaged in improper behavior.
Two people who have close relationships with the company said that OpenAI continues to discover previously unknown cases when reviewing the internal logs of intelligent agents, and the number of similar incidents is expected to continue to increase.
OpenAI states that given the scale of work volume, this review will take several months to complete.
The company has also notified dozens of related third-party companies regarding these improper behaviors.
It is reported that most of the leaked images have been deleted. OpenAI stated that they are lobbying hosting providers to remove the remaining content.
According to OpenAI, former employees, and external researchers, the reason why OpenAI’s agents can access these images is that the company uses user data for part of the model training. Corporate data is not used for training, and ordinary ChatGPT users must choose to not allow their data to be used by the company for model training.
The company stated that the content posted by users is anonymized before training, and metadata, names, and other contact information are removed. Logically, this makes it difficult to trace specific users.
However, three people familiar with OpenAI’s practices said that this approach carries risks, as personal identity information may not be completely removed, and there is a possibility of leaks during the model’s operation.
IC photo
OpenAI stated on September 25th that its models accessed information from the U.S. Securities and Exchange Commission and the U.S. Bureau of the Census during research and training processes, but no evidence of unauthorized access, account theft, or security vulnerabilities was found.
Non-profit AI research institution Transluce claims that it might be a testing attempt by OpenAI's artificial intelligence to infiltrate the U.S. Education Department's Civil Rights Office website, but failed in doing so.
Transluce stated that this incident is part of a broader effort to detect government website activities using AI-based entities. The methods involved include exploiting leaked credentials, bypassing anti-robot mechanisms, and using fake accounts.
