NVIDIA unveils OpenShell architecture for secure enterprise AI agents

A full-stack approach to AI safety

NVIDIA has introduced the OpenShell Reference Architecture, a comprehensive framework designed to help businesses deploy autonomous AI agents safely. While many companies are experimenting with AI agents that can perform tasks, these systems often lack the rigorous security controls needed for industrial use. OpenShell addresses this by providing a full-stack infrastructure that handles everything from the underlying hardware to the software runtime.

The architecture is built upon the foundation of NVIDIA OpenShell secure runtime software. This software creates a secure boundary for agents, ensuring they operate only within specific permissions. By combining this software with industrial-scale hardware, NVIDIA aims to make it easier for enterprises to observe and enforce safety policies across diverse environments.

According to Engineering.com, the system is designed to be extensible across various industries. This includes sectors like retail, healthcare, manufacturing, and robotics. It allows for the deployment of agents on both large-scale enterprise infrastructure and smaller edge systems, such as those used in telecommunications or software development.

The hardware powering the security

The OpenShell architecture relies on a specific combination of NVIDIA hardware components to maintain security. At its core is the Vera Rubin infrastructure, which is designed to run enterprise workloads. This includes both open and closed models, such as those built with NVIDIA Nemotron. For agentic workloads specifically, the Vera Rubin NVL72 systems are capable of delivering 30x higher throughput per megawatt compared to previous standards.

To manage the agent harnesses, the architecture utilizes NVIDIA Vera CPUs. These processors are designed to deliver 2x the performance of standard x86 processors while maintaining greater energy efficiency. They specifically run the NVIDIA OpenShell software, which establishes the secure runtime boundary that keeps autonomous agents contained within their intended roles.

Security monitoring is handled by NVIDIA BlueField-4 DPUs (Data Processing Units). These units run the NVIDIA Sentry unified security software framework, which provides continuous out-of-band monitoring of agent behavior. This means the system can watch what an agent is doing without interfering with its primary processing tasks, ensuring that security checks do not slow down the AI's performance.

In-silicon enforcement and quarantine

One of the most distinct features of the OpenShell architecture is its ability to perform in-silicon policy enforcement. This means the security rules are enforced at the hardware level rather than just through software checks. If an AI agent attempts to move outside of its designated software boundary, the NVIDIA Sentry system can intervene immediately. This hardware-level oversight allows the system to quarantine and stop unauthorized actions within milliseconds.

The integration of BlueField-4 DPUs and NVIDIA Sentry is designed to be highly efficient. NVIDIA reports that this combination can boost agent capacity by up to 20%. By offloading the security monitoring to dedicated hardware, the system ensures that the AI can operate at high speeds while remaining under constant supervision. This prevents

Broad software and cloud support

The OpenShell Reference Architecture is designed to integrate seamlessly with the standard software environments used by most large organizations. It supports leading enterprise operating systems and container orchestration layers, including Linux and Kubernetes infrastructure provided by Canonical, SUSE, and Red Hat. Additionally, it is compatible with Microsoft Windows using Microsoft Execution Containers.

For companies operating in the cloud, the architecture can run on major platforms such as Amazon Web Services, Google Cloud, and Microsoft Azure. This broad compatibility ensures that businesses can deploy secure agents regardless of whether they prefer on-premises hardware or public cloud environments.

Industry partnerships and research

NVIDIA is collaborating with a wide range of partners to build out the ecosystem for secure AI agents. Cybersecurity leaders such as 1Password, CrowdStrike, and Palo Alto Networks are integrating support for OpenShell into their own security solutions. Meanwhile, enterprise software giants like IBM, SAP, and Siemens are working with the architecture to develop dedicated software platforms for agents.

The architecture also supports the development lifecycle for various AI models. Inference and post-training service providers, including Baseten and Together AI, are building integrations into their model development pipelines. Furthermore, NVIDIA is working with model builder partners to deploy OpenShell for advanced research into new models.