NVIDIA Puts Agent Safety in Silicon With the Open Agent Safety Platform

NVIDIA announced the Open Agent Safety Platform on 28 September 2026, a reference system design and set of software tools for governing AI agents across hardware, compute and robotics.

The distinguishing choice is where the boundary sits. Most agent-safety work happens in the application layer, where the thing being constrained is also the thing doing the constraining. NVIDIA push it downwards: OpenShell, a secure runtime, establishes limits for agents at the system level, and the platform reaches into the Vera CPU and the BlueField-4 DPU — a data processing unit that sits between the agent and the network it wants to reach.

That matters most for agents that touch the physical world. An agent writing code can be sandboxed in software; an agent driving a robot needs a boundary that holds when the software misbehaves.

The platform covers the full span from testing through to deployment, which is NVIDIA’s framing rather than a claim this page can evaluate — the design is newly published and there is no independent assessment of it yet.