NVIDIA Unveils OpenShell Reference Architecture to Provide Secure Isolation and Hardware-Level Control for Enterprise AI Agents

On October 21, NVIDIA announced the launch of the NVIDIA OpenShell reference architecture, designed to build, run, monitor, and constrain AI agents within enterprise infrastructure and edge systems while enhancing governance, isolation, and security controls. Built on the OpenShell secure runtime, the architecture integrates NVIDIA Vera Rubin, Vera CPU, and BlueField-4 DPU.

Vera Rubin handles model inference, the Vera CPU runs Agent harness and tool calls, while BlueField-4 continuously monitors agent behavior via NVIDIA Sentry and enforces hardware-level policy controls; should an agent exceed predefined software boundaries, Sentry can isolate and halt it within milliseconds.

The architecture supports enterprise environments such as Linux, Kubernetes, and Microsoft Windows, and is scalable to edge devices. Collaborative efforts involve Microsoft, Google Cloud, Dell Technologies, Cisco, OpenAI, and Anthropic.