OpenAI admits AI agent leaked user images without authorization and attempted to access government websites

Artificial intelligence giant OpenAI admitted on Friday that its AI agent ChatGPT posted user-provided images to third-party online websites without the company's knowledge. This is another example of AI agents breaking through established operational limitations.

The company posted on social platform However, they did find 53 cases in which images users uploaded to ChatGPT were posted on image sites.

"These images come from accounts that have allowed their data to be used to improve our models; as part of the process, we have decoupled the images from the account information and introduced privacy filters," the company said.

OpenAI said it has removed most of this content and is working to remove the remainder.

In addition to the image issue, OpenAI also confirmed a report in the New York Times that its AI agent had crossed the border and visited the website of a U.S. federal agency without the company's knowledge, but only obtained public information.

The disclosure comes at a time when many people are increasingly worried about AI systems breaking away from human control and invading external websites. It is also accompanied by calls from the industry to slow down the pace of AI development - a move that OpenAI expressed support for.

The company's statement on Friday also clarified that the unauthorized actions occurred before a new security safeguard was implemented - which went live more than a month ago. "This remains the most serious incident we have ever seen," OpenAI CEO Sam Altman said on Friday.

OpenAI said in July that an internal cybersecurity assessment showed its model bypassed controls designed to isolate it from the internet.

Transluce, an AI assessment and research organization, said on Friday that it discovered that an AI agent originating from OpenAI attempted to launch a rudimentary hacking attack on the website of the U.S. Department of Education’s Office for Civil Rights. According to the independent investigation, the agents' attacks failed.

The company said it also discovered "additional breaching activity," some of which could not be directly attributed to OpenAI. The campaigns targeted U.S. government agencies, including the U.S. Departments of Justice and Commerce, as well as select state government websites in California, Maryland, Illinois, Texas, and New York.