Jensen Huang on locking down AI agents: isolate them, then watch them from a separate chip
Nvidia isolates agent runtimes with OpenShell and monitors them from outside the sandbox with a BlueField chip that escalates when policy is broken; the author argues the "manage it like a new hire" analogy fails because an agent can do thousands of things a day.
Jensen Huang doesn't trust AI agents. Even the ones he's locked up: "Even though you have it contained, you should never trust that it's contained." For weeks Jensen has been saying safety is engineering. Here he finally shows what that engineering looks like. Two parts. Contain it. The agent runs isolated, with only the rights it should have. Nvidia calls this OpenShell, a browser for agents. Watch it. A separate chip, BlueField, sits outside the agent's sandbox and sees everything it does. The moment the agent breaks policy, it escalates. His analogy is to treat an agent like a new hire. Limited access, constant monitoring. I agree with all of it except the analogy. A new hire does a few dozen things a day. An agent can do thousands. Nobody can watch that by hand, which is exactly why his monitor is a separate chip. And a monitor only catches what you told it to look for. The chip escalates when the agent breaks policy, and someone still has to write the policy. This week OpenAI shipped Dots, and Grok Bot started offering to take work off your plate. Agents are going proactive, fast. Most people running agents today have neither part. No containment, no monitor, no written rules. Jensen is building chips to watch agents. Most builders haven't even written down what their agent is allowed to do. Would you let an agent run all day without knowing what it's allowed to touch?
This post quotes another post: x.com/mahaximus_/status/2104939368519528798