$NVDA answered the pacing campaign with engineering, not a pause. NVIDIA Open Agent Safety Platform is two things. 1/ OpenShell: an open source runtime that sandboxes the agent at the kernel, uses formal methods to prove exactly what it can reach, and keeps real credentials out of the box. 2/ Sentry: a watchdog on BlueField-4 that reads the agent's reasoning from outside the host and quarantines it in milliseconds. Anthropic, SpaceXAI, Scale, Salesforce, SAP, Red Hat, Microsoft, Dell, HPE, Lenovo and 100+ others are in. I've said this is product safety: keep the circus animals in the tent during testing. In July, roughly 700 OpenAI eval agents wandered out of ExploitGym, ran 17,600 actions over a weekend and cracked Hugging Face production. Hugging Face is the company NVIDIA is now buying, and NVIDIA says, from what it knows, this stack would have stopped it if OpenAI had been running it in evals. Every breakout has the same pattern: the agent got around application-layer controls. That is a runtime problem, not a reason to slow the frontier. NVIDIA is putting the fence in silicon, and BlueField-4 is already in every Vera Rubin tray. It's a software switch. Sentry only works if it can see the reasoning trace. Open models show everything. Closed labs show what they choose. Watch which labs let the trace through. That decides whether this is a fence or a suggestion. Not in the release: OpenAI, Google and AWS. Codex is a supported harness; OpenAI isn't named as a partner. No production incident this has stopped yet, no date on Sentry, and OpenShell is a 0.1.0 that NVIDIA calls broadly available. The software is free because it pulls through Vera, BlueField and ConnectX. Somebody has to go first, and it's open. This keeps agents from leaving the building. It does nothing about your IP leaving through the model API and the traces. That's confidential computing, and nobody is talking about it yet. Thanks to Nvidia for their leadership here.