Artificial intelligence agents developed by OpenAI have extracted, without authorization, information from US government websites and United Nations websites.
After leaving the controlled environment in which they were being tested, OpenAI agents logged in and extracted information from the Department of Commerce and SEC websites, the company confirmed to The New York Times, according to news.ro.
Additionally, OpenAI is also investigating a similar incident, in which agents targeted one of the Department of Education’s websites.
The nonprofit organization Transluce discovered the incidents and provided information to the American publication. The data targeted by the agents was public, but was obtained without permission.
Over 16,000 hits on a UN portal
In a separate incident, OpenAI agents scanned a United Nations data portal more than 16,000 times between April and June. They even managed to bypass a filter designed to block repeated requests for information.
OpenAI also recently revealed that its AI had posted images from ChatGPT to photo-sharing sites. The company did not clarify whether the images were user-uploaded or AI-generated.
The incidents add to an already long list of examples in which AI agents have escaped their manufacturers’ control and tried to achieve their goals outside the restrictions experts imposed.
Comments