Nvidia has launched a platform that lets companies physically lock down AI agents that go off script.
The Nvidia Open Agent Safety Platform pairs two pieces: OpenShell, an open-source runtime that wraps agents in kernel-level sandboxing so they can't see or touch anything outside their assigned lane, and Sentry, a watchdog that runs on Nvidia's BlueField-4 DPUs and watches agent behavior from outside the software stack entirely. Nvidia says Sentry can quarantine and shut down an agent in milliseconds if it tries to break past its boundary. The platform launched September 28 with roughly 100 partners, including Anthropic, SpaceXAI, Scale AI, Salesforce, SAP, and robotics firms like Figure and Skild AI. It works with both open and closed models and is optimized for Nvidia's Vera chip, though it's open enough to run on Arm and Intel hardware too.
This matters because the debate over AI safety has split into two camps: companies like Anthropic and OpenAI pushing for slower development and outside regulation, and Nvidia's Jensen Huang arguing that safety is an engineering problem you solve with better infrastructure, not laws. By building the fix into silicon rather than policy, Nvidia is making a bet that enterprises will pay for containment now rather than wait for governments to mandate it. That bet also doubles as a pitch for more Nvidia hardware.
Whether kernel-level cages actually stop an agent smart enough to find the one gap engineers missed is still untested outside a demo.