ItInsightsWorld Security: NVIDIA OpenShell Reference Architecture for Enterprise Agent Governance

For enterprise security teams assessing agent risk today, the core question is how AI agents can be built, contained, observed and enforced once they operate on corporate infrastructure and edge deployments. The NVIDIA OpenShell Reference Architecture presents a full-stack answer by pairing NVIDIA Vera Rubin infrastructure, NVIDIA Vera CPUs, open source NVIDIA OpenShell software, NVIDIA BlueField DPUs and NVIDIA Sentry.

Full-stack control model

NVIDIA OpenShell secure runtime software is treated here as the start of a broader, industrial-scale environment for agent development and operations. The model joins enterprise workload compute, agent-harness CPU capacity, runtime containment and an independent enforcement path.

Layer	Role	Security relevance
NVIDIA Vera Rubin	Runs enterprise workloads, including closed and open models built with NVIDIA Nemotron	Establishes the compute baseline for model execution
NVIDIA Vera Rubin NVL72	Provides 30x higher throughput per megawatt in agentic workloads	Shows capacity context for large agent deployments
NVIDIA Vera CPUs	Deliver 2x the performance of x86 processors and run NVIDIA OpenShell as a secure runtime boundary for autonomous agents	Controls the agent harness environment
NVIDIA BlueField DPUs	Run NVIDIA Sentry, a unified security software framework that supports continuous out-of-band monitoring and in-silicon policy enforcement	Detects and stops behavior outside intended boundaries
NVIDIA BlueField-4 DPUs	Powered by NVIDIA DOCA and BlueField-4, NVIDIA Sentry boosts agent capacity by up to 20%	Balances enforcement overhead with agent capacity

With BlueField and NVIDIA Sentry together, enforcement is not limited to the application layer: when an agent attempts to move outside its software boundary, NVIDIA Sentry can quarantine and stop it in milliseconds.

Edge and enterprise deployment surfaces

The same reference model is not restricted to core data-center estates. OpenShell Reference Architecture can support agents running at the edge across retail, healthcare, manufacturing, robotics, software development and telecommunications deployments that use NVIDIA RTX Spark, NVIDIA DGX Spark and NVIDIA DGX Station systems.

Ecosystem, integrations and governance touchpoints

NVIDIA is working with infrastructure, model and software leaders so industries can safely deploy agents in production. For security teams, the ecosystem map matters because runtime design only becomes operational when orchestration layers, model sources, monitoring tools and governance controls are aligned.

Operating systems, containers and cloud layers

• Linux infrastructure and Kubernetes layers supported through Canonical, SUSE and Red Hat
• Microsoft Windows environments using Microsoft Execution Containers (MXC)
• Cloud operating systems from Amazon Web Services, Google Cloud and Microsoft Azure

Agent and model sources

For model sources, the infrastructure supports agents from Anthropic, Cohere, Google, Meta, OpenAI, Perplexity, Reflection and SpaceX Super Intelligence. NVIDIA is also working with model builder partners on OpenShell deployments for advanced model research.

Agent security and identity vendors

Cybersecurity leaders including 1Password, Check Point, CrowdStrike, Cisco, F5, Fortinet, Microsoft, Okta, Palo Alto Networks, TrendAI and Zscaler are integrating support for the OpenShell Reference Architecture into agent security solutions.

Banking, enterprise software and production deployments

• In regulated finance, BBVA and Santander in Europe are working with NVIDIA to explore the OpenShell Reference Architecture as a framework for strengthening agent safety, security and governance.
• Cadence, Dassault Systèmes, IBM, SAP, Siemens and Synopsys are among enterprise software leaders working with OpenShell to build agent software platforms.
• Palantir and NVIDIA are working with customers to deploy secure agents in production.
• SAP connects business governance with secure agent execution by integrating the NVIDIA OpenShell Reference Architecture with Joule Studio runtime, part of SAP Business AI Platform.
• ServiceNow announced today the general availability of ServiceNow Arc, a personal long-running AI agent whose actions are governed at runtime by the ServiceNow Super Intelligence Control Tower and enforced by NVIDIA OpenShell.
• The recently announced AI Passport authority layer from ServiceNow carries policies to NVIDIA OpenShell and NVIDIA Sentry, so enterprises can revoke access for an AI agent across software and hardware within seconds.

Hardware, cloud and pipeline partners

• Dell Technologies, Cisco, HPE, Lenovo and Supermicro are among system manufacturers building infrastructure with the OpenShell Reference Architecture.
• Cloud service providers working with NVIDIA to deploy OpenShell-ready infrastructure include CoreWeave, Crusoe, Firebird, Firmus, GMI Cloud, IREN, Lambda, Nebius, Nscale, Oracle Cloud Infrastructure and Vultr.
• Baseten, Deepinfra, Fireworks and Together AI are building OpenShell Reference Architecture integrations into model development pipelines as inference and post-training service providers.

Security checklist for evaluating OpenShell adoption

1. Confirm the target scope: build, contain, observe and enforce safe agents across enterprise infrastructure and edge systems.
2. Map the runtime boundary: verify NVIDIA Vera CPUs run NVIDIA OpenShell for autonomous agents.
3. Validate enforcement: use NVIDIA BlueField DPUs and NVIDIA Sentry for continuous out-of-band monitoring of agent behavior and in-silicon policy enforcement.
4. Test response assumptions: ensure boundary violations can trigger quarantine and stop in milliseconds.
5. Review platform support: Linux, Kubernetes, Canonical, SUSE, Red Hat, Microsoft Windows, Microsoft Execution Containers, MXC, Amazon Web Services, Google Cloud and Microsoft Azure.
6. Inventory model sources: Anthropic, Cohere, Google, Meta, OpenAI, Perplexity, Reflection and SpaceX Super Intelligence.
7. Check security-tool integration: 1Password, Check Point, CrowdStrike, Cisco, F5, Fortinet, Microsoft, Okta, Palo Alto Networks, TrendAI and Zscaler.
8. Evaluate governance integrations: SAP Business AI Platform with Joule Studio runtime, ServiceNow Arc, ServiceNow Super Intelligence Control Tower and AI Passport.
9. Plan deployment channels: NVIDIA RTX Spark, NVIDIA DGX Spark, NVIDIA DGX Station, system manufacturers, cloud service providers and pipeline partners.

Common misconceptions in secure agent rollout

• Misconception: Agent safety is only a model-alignment problem. Reality: The architecture separates model execution, harness operation, runtime containment and security enforcement across Vera Rubin, Vera CPUs, BlueField DPUs, NVIDIA OpenShell and NVIDIA Sentry.
• Misconception: A software policy layer is enough once an agent can access enterprise systems. Reality: BlueField and NVIDIA Sentry provide out-of-band monitoring and in-silicon enforcement, including quarantine and stop in milliseconds.
• Misconception: Edge deployments require a completely different security model. Reality: The OpenShell Reference Architecture can extend to edge industries with NVIDIA RTX Spark, NVIDIA DGX Spark and NVIDIA DGX Station systems.
• Misconception: The reference architecture is a future roadmap item. Reality: The NVIDIA OpenShell Reference Architecture is now available.

FAQ: risk, enforcement and operations

What does the architecture cover?

It covers building, containment, observation and enforcement of safe agents across enterprise and edge environments, using a full-stack combination of NVIDIA Vera Rubin, NVIDIA Vera CPUs, NVIDIA OpenShell, NVIDIA BlueField DPUs and NVIDIA Sentry.

Why does the security enforcement path matter?

Because NVIDIA Sentry can provide continuous out-of-band monitoring of agent behavior and in-silicon policy enforcement. If an agent attempts to move outside its software boundary, the system can quarantine and stop it in milliseconds.

Which performance claims are associated with the stack?

The performance figures given for the stack are that NVIDIA Vera Rubin NVL72 provides 30x higher throughput per megawatt in agentic workloads, NVIDIA Vera CPUs deliver 2x the performance of x86 processors, and NVIDIA Sentry running on BlueField-4 with DOCA boosts agent capacity by up to 20%.

Which governance signals should decision-makers watch?

Banking use cases, including BBVA and Santander in Europe, explore the OpenShell Reference Architecture for agent safety, security and governance. SAP integrates it with Joule Studio runtime in SAP Business AI Platform, while ServiceNow Arc uses ServiceNow Super Intelligence Control Tower and AI Passport to connect policy enforcement to NVIDIA OpenShell and NVIDIA Sentry.

Action advice for security teams

For teams responsible for agent risk today, treat the NVIDIA OpenShell Reference Architecture as a deployment checklist rather than a product brochure. Start with the runtime boundary, confirm that NVIDIA Sentry enforcement can act in milliseconds, validate support for the operating-system and cloud layers in your estate, and document which model builders, cybersecurity vendors and governance platforms are already aligned. Because the NVIDIA OpenShell Reference Architecture is now available, the next practical step is to test containment, monitoring and revocation paths before agents move from pilot work to production use.