NVIDIA Introduces OpenShell Reference Architecture for Isolated AI Agent Deployments
NVIDIA has released a reference architecture and technical documentation for controlled AI agent workspaces. The design combines an isolated runtime, infrastructure and DPU-level security policy enforcement.
NVIDIA OpenShell architecture is a new reference design for controlled and isolated AI agent deployments. NVIDIA announced it on October 21, 2026, along with accompanying technical documentation focused on secure agent workspaces that access internal tools, data and APIs.
It is not a standalone universal security standard, but a design for combining an agent runtime, computing infrastructure and policy enforcement mechanisms. NVIDIA wants to move some control from the application layer into the isolated runtime, network and DPU layers.
NVIDIA OpenShell architecture connects the runtime and DPU
The reference design is intended to connect the NVIDIA OpenShell runtime with Vera Rubin and Vera CPU computing infrastructure. Security enforcement is to be provided by the BlueField-4 DPU together with the NVIDIA Sentry platform.
The purpose of this arrangement is to create a workspace in which agent policies are not enforced only within the agent’s application. For enterprise deployments, this approach may be particularly relevant for agents that perform tasks through corporate systems or work with sensitive data.
NVIDIA states that it supports enterprise Linux/Kubernetes environments and Microsoft Windows with Microsoft Execution Containers. The published documentation also includes reference implementations for Red Hat OpenShift and Microsoft Azure.
ServiceNow confirmed use with the Arc agent
ServiceNow independently confirmed an expansion of its collaboration with NVIDIA. Its Arc agent uses NVIDIA OpenShell for runtime policy enforcement, meaning rules applied while the agent is running.
ServiceNow also distinguishes confirmed use from planned features. The company describes AI Passport with enforcement through OpenShell and BlueField as a future release. The announcement therefore does not mean that all announced capabilities are already available in production environments.
Vendor performance claims await independent testing
In its announcement, NVIDIA states that the agent can be stopped in milliseconds, with 30 times the throughput per megawatt and an increase in agent capacity of up to 20%. These are vendor product claims that have not been independently verified.
Likewise, the involvement of partners does not mean that all of them are already operating the solution in production. Some integrations and hardware features may be under development or depend on the future availability of specific components.
The architecture may reduce risks when operating more autonomous agents, but by itself it does not guarantee their security or completely prevent misuse. The specific implementation, the customer’s policy settings and their integration with the systems in use will be decisive.
What to watch next
- the actual availability and licensing terms of the individual OpenShell components and related hardware,
- independent tests of isolation, monitoring and responses to policy violations,
- the timelines for ServiceNow and other partners to make the announced integrations available for production use.
Sources
- NVIDIA Newsroom – Confirms the announcement of the OpenShell Reference Architecture, its stated components, partners and documentation availability.
- NVIDIA Documentation – Confirms the published technical documentation, security model and reference implementations for OpenShift and Azure.
- ServiceNow Newsroom – Independently confirms OpenShell integration with the Arc agent and distinguishes available features from future AI Passport features.
Verified and updated: October 21, 2026 15:21
