NVIDIA Builds a Full-Stack Safety System for AI Agents

NVIDIA has announced the NVIDIA OpenShell Reference Architecture, a full-stack approach for building, containing, observing and enforcing safe AI agents across enterprise infrastructure and edge systems. The reference architecture is now available and builds on NVIDIA OpenShell secure runtime software.

The goal is to place safety and control inside the computing architecture rather than treat them as separate software features. The design brings together systems for model inference, agent operations and hardware-level policy enforcement, creating defined boundaries for agents as they work across enterprise environments.

Three layers for thinking, acting and control

At the center of the architecture is NVIDIA OpenShell, which provides a secure runtime boundary for autonomous agents. NVIDIA Vera Rubin infrastructure handles enterprise workloads, including closed and open models built with NVIDIA Nemotron, while NVIDIA Vera CPUs operate agent harnesses and tool use.

NVIDIA BlueField-4 DPUs run NVIDIA Sentry, a unified security software framework built on NVIDIA DOCA software. Sentry provides continuous out-of-band monitoring of agent behavior and in-silicon policy enforcement, allowing security controls to operate alongside the agent rather than only through its software environment.

The system is designed to stop an agent that moves outside its software boundary. When that happens, NVIDIA Sentry quarantines and stops the agent in milliseconds. NVIDIA says Sentry, powered by DOCA and running on BlueField-4, can boost agent capacity by up to 20%.

Jensen Huang, founder and CEO of NVIDIA, described the structure in simple terms: “AI capability and safety must advance together. As enterprises prepare to deploy billions of AI agents, safety and control must be built into the computing architecture.”

Huang added: “NVIDIA created OpenShell to give agents the freedom to do useful work within clearly defined boundaries — contained, monitored and under control. With the OpenShell Reference Architecture, agents can think on Rubin, act on Vera and be governed on BlueField.”

Performance across enterprise and edge systems

The architecture pairs NVIDIA accelerated computing with software for secure agent operations. NVIDIA Vera Rubin, with NVIDIA Vera Rubin NVL72 infrastructure at its core, is fungible across pretraining, post-training and inference. For agentic workloads, NVIDIA says Vera Rubin NVL72 systems deliver 30x higher throughput per megawatt.

NVIDIA Vera CPUs deliver 2x the performance of x86 processors with greater energy efficiency, and they run NVIDIA OpenShell. That gives the agent harnesses a place to operate while the broader architecture connects model workloads, tools and security enforcement.

The design also extends beyond data centers. NVIDIA RTX Spark, NVIDIA DGX Spark and NVIDIA DGX Station systems support agents at the edge across retail, healthcare, manufacturing, robotics, software development and telecommunications.

OpenShell runs on leading enterprise operating systems and container orchestration layers, including Linux and Kubernetes infrastructure powered by Canonical, SUSE and Red Hat. It also supports Microsoft Windows through Microsoft Execution Containers, known as MXC, and can run on cloud operating systems from Amazon Web Services, Google Cloud and Microsoft Azure.

A broad network of hardware, software and security partners

NVIDIA says it is working with companies across models, infrastructure, cybersecurity and enterprise software to support production deployments. Enterprises can run models from Anthropic, Cohere, Google, Meta, OpenAI, Perplexity, Reflection and SpaceX Super Intelligence on OpenShell infrastructure, while NVIDIA is also working with model builders on advanced model research.

Cybersecurity companies integrating support for the architecture include 1Password, Check Point, CrowdStrike, Cisco, F5, Fortinet, Microsoft, Okta, Palo Alto Networks, TrendAI and Zscaler. System manufacturers Dell Technologies, Cisco, HPE, Lenovo and Supermicro are building infrastructure with the OpenShell Reference Architecture.

Cloud service providers working 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 integrations for inference and post-training service pipelines.

The architecture also reaches into enterprise applications. Cadence, Dassault Systèmes, IBM, SAP, Siemens and Synopsys are working with OpenShell to build agent software platforms, while Palantir and NVIDIA are working with customers to deploy secure agents in production.

SAP is integrating the architecture with Joule Studio runtime, part of SAP Business AI Platform. ServiceNow has announced 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.

ServiceNow’s AI Passport authority layer carries those policies to NVIDIA OpenShell and NVIDIA Sentry, allowing enterprises to revoke an AI agent’s access across software and hardware within seconds. Leading banks, including BBVA and Santander, are also working with NVIDIA to explore the architecture as a framework for agent safety, security and governance.

The partner list includes Anthropic, AWS, Cisco, CrowdStrike, Dassault Systèmes, Dell Technologies, Google Cloud, HPE, IBM, Meta, Microsoft, OpenAI, Palo Alto Networks, Perplexity, SAP, ServiceNow, Siemens and SpaceX Super Intelligence, among others. NVIDIA’s reference architecture is now available.

Stay connected via Google News

Follow us for the latest AI News