NVIDIA launches OpenShell Reference Architecture — full-stack, in-silicon governance for safe AI agents

At GTC Berlin (Oct 21), NVIDIA announced the OpenShell Reference Architecture: an open hardware+software stack for building, containing, observing and enforcing safe agents, built on the OpenShell secure runtime, Vera Rubin inference, Vera CPUs for agent harnesses, and BlueField-4 DPUs running NVIDIA Sentry for out-of-band monitoring and in-silicon policy enforcement (quarantine in milliseconds). Anthropic, AWS, Cisco, CrowdStrike, Google Cloud, IBM, Microsoft, OpenAI, Palo Alto, ServiceNow and others are integrating; ServiceNow Arc GA is enforced by OpenShell, and 1Password, Check Point, F5, Fortinet, Okta, TrendAI and Zscaler are adding support.

This is the first silicon-level agent security architecture from a major infrastructure vendor, moving agent governance from software policy to hardware-enforced runtime containment — a reference point for every agent-security product roadmap this cycle.

Enterprises deploying agents at scale and security vendors building agent protections should evaluate OpenShell as an enforcement substrate; system integrators and AI-factory builders should track Vera CPU/BlueField requirements before committing to agent-runtime architectures.