Skip to content
Neural Daily

ToolsNews

NVIDIA OpenShell reference design secures AI agents

At GTC Berlin, NVIDIA released the OpenShell Reference Architecture, a full-stack blueprint for running agents inside contained, monitored boundaries. Policies are enforced in hardware by Sentry on BlueField DPUs.

By: Mikola Máté Patrik · · 2 min read

NVIDIA press image: NVIDIA OpenShell Reference Architecture
Image: NVIDIA · Press kit · Official NVIDIA press/blog image · Source

What changed?

On October 21 at GTC Berlin, NVIDIA announced the OpenShell Reference Architecture (NVIDIA Newsroom). It builds on the open source OpenShell runtime, which uses kernel-level sandboxing and policies to sit between an agent and the systems it can touch. The architecture is available now.

Three layers

  • Vera Rubin NVL72 handles model inference, for closed models or open Nemotron-based ones.
  • Vera CPUs run the agent harness and tool use inside OpenShell sandboxes.
  • BlueField-4 DPUs run NVIDIA Sentry, which monitors agent behavior out of band. According to NVIDIA, if an agent steps outside its boundary, Sentry quarantines it within milliseconds.

At the edge, RTX Spark, DGX Spark and DGX Station systems extend the same design. Supported environments include Linux and Kubernetes from Canonical, SUSE and Red Hat, Windows via Microsoft Execution Containers (MXC), plus AWS, Google Cloud and Azure.

Who is on board?

The partner list is long. Models from Anthropic, OpenAI, Meta, Google, Cohere and Perplexity can run on OpenShell infrastructure. CrowdStrike, Palo Alto Networks, Cisco, Okta, Zscaler and others are adding support to their security products. SAP is integrating it with the Joule Studio runtime. ServiceNow made Arc, its long-running personal agent, generally available today, with actions enforced by OpenShell. In Europe, BBVA and Santander are exploring it.

Alongside: Confidential Bare Metal

The same day, NVIDIA expanded Confidential Computing (NVIDIA blog). The new Confidential Bare Metal option lets frontier model builders keep control of their models and data on third-party bare-metal infrastructure. Support now covers Grace Blackwell NVL72 rack-scale systems and HGX and DGX Rubin NVL8. Cohere plans to use it in Model Vault, and Google Cloud is offering it on A4X Max instances.

Why it matters

Teams taking agents to production have mostly stitched permissions together from software guardrails. NVIDIA now offers a reference that reaches down to the hardware. Performance claims, such as Sentry boosting agent capacity by up to 20%, are vendor data with no independent verification yet. Practical next step: check whether your cloud or security vendor is on the partner list.

    Produced with the help of AI, reviewed by an editor. How we make our articles

    Reviewed and edited by the Neural Daily desk. About

    Related articles