Quoting Victoria Kim

OpenAI has added monitoring that lets staff immediately intervene to halt training if its models access the internet in unauthorized ways.