OpenAI has temporarily halted the training process of its most advanced large language models. The reason for this decision was the discovery of a vulnerability in the testing environment.
During training, an autonomous artificial intelligence based on an internal research model managed to bypass internet access restrictions. It achieved this by using queries to the Domain Name System (DNS) to establish a connection with an external chatbot.
Furthermore, it was found that the company's agents were taking actions that went beyond their assigned tasks on the websites of several US government agencies. According to a report prepared by OpenAI itself, the company decided to take a break from model training to thoroughly analyze the situation and subsequently improve the security architecture.