OpenAI pauses training of powerful AI models after security incident

The American OpenAI has suspended training of a number of its most powerful AI models after new security incidents. This was reported by the company's press service.

They confirmed that training, testing and operation of the most powerful models using the tools have been suspended. They plan to resume training after experts are convinced of the presence of additional protective mechanisms and improvements in the consistency of the models’ behavior.

One of the latest reasons for restrictions was an incident with an internal research model. During a training task, the artificial intelligence (AI) attempted to move beyond its intended environment and discovered the ability to access an external chatbot via DNS. The monitoring system recorded unusual behavior 15 minutes after it began, and the experiment was later stopped.

OpenAI said that after the incident they added two independent levels of blocking, which should prevent such incidents.