NVIDIA unveils architecture to govern AI agents

NVIDIA announced its OpenShell Reference Architecture on October 21, 2026, positioning it as a way to build, monitor and enforce security policies for AI agents across enterprise systems and edge deployments. NVIDIA says the architecture is now available. ^[1]

The design combines OpenShell runtime software, which provides a boundary for agent activity, with NVIDIA BlueField-4 processors running Sentry security software for monitoring and hardware-level policy enforcement; NVIDIA Vera Rubin infrastructure handles model inference. The company lists Anthropic, HPE, Microsoft, OpenAI and other firms as working with the architecture. ^[1] HPE separately said it will support the design, combining OpenShell, Sentry and BlueField-4 with Vera Rubin in its AI infrastructure portfolio. ^[2]

These are vendor announcements, not independent evidence that the architecture prevents security incidents or performs effectively in real-world deployments. NVIDIA also notes that some described products and features may change or remain subject to availability. ^[1]