In just 35 seconds, OpenAI can shut down its AI models and agents using its new control system. Even if the system itself fails, active agents will stop running automatically. The mechanism should help prevent incidents like the hug attack. A similar "switch" already works in user chats.