Nvidia releases OpenShell Reference Architecture for safe AI agent execution

Nvidia announced the OpenShell Reference Architecture at its GTC Berlin conference on Wednesday, introducing a comprehensive design for running artificial intelligence agents within strict operational limits. This new framework extends from large-scale enterprise infrastructure down to smaller edge computing machines, aiming to provide consistent safety controls across diverse deployment environments.

The design builds upon OpenShell, an open-source runtime software previously released under the Apache 2.0 licence, which establishes a defined boundary for agent actions. According to Nvidia, this architecture is part of a broader agent safety platform launched in September, ensuring that models running on Vera Rubin systems operate within secure parameters while driver software executes on Vera CPUs.

Technical safeguards and monitoring

Central to this architecture is Nvidia Sentry, software that runs on BlueField-4 data processing units to observe agent behavior from outside the main system. Nvidia stated that if an agent attempts to move beyond its designated software boundary, Sentry can quarantine and stop the process in milliseconds. This external monitoring approach reportedly increases agent capacity by up to 20%, allowing organizations to maintain higher performance levels without compromising on safety oversight.

The reference architecture also addresses edge computing scenarios, covering systems such as RTX Spark, DGX Spark and DGX Station. Nvidia recently demonstrated RTX Spark laptops designed to run these agents locally, illustrating how the safety constraints apply even to decentralized, on-device processing units that operate away from central data centers.

Industry adoption and partnerships

Nvidia indicated that a wide range of AI developers are working with this infrastructure, including Anthropic, Cohere, Google, Meta, OpenAI, Perplexity, Reflection and SpaceX Super Intelligence. Security vendors such as 1Password, Check Point, CrowdStrike, Cisco, Fortinet, Okta, Palo Alto Networks and Zscaler are adding support for the platform, integrating their tools into this safety ecosystem.

Anthropic has partnered with Nvidia since September, pairing its Claude Managed Agents service with OpenShell to allow companies to limit agent actions and log every step. Financial institutions like BBVA and Santander are exploring the technology to strengthen governance, while SAP is building it into Joule Studio and ServiceNow has enforced its Arc agent actions with OpenShell.

What should you do?

Organizations preparing to deploy AI agents should review the OpenShell Reference Architecture to understand how fixed limits can be applied to their specific infrastructure. Security teams should evaluate whether current agent deployments operate within defined boundaries or if they require additional monitoring mechanisms like Nvidia Sentry. IT leaders should assess compatibility with existing security vendor tools, as many major providers are adding support for this framework.

Background: Edge computing

Edge computing processes data near its source instead of sending it to a distant data center. This approach lowers delay, saves bandwidth, and keeps sensitive information local. It shifts security responsibilities from the core network to individual devices.

Read the full guide: Edge Computing Explained: Why Processing Data Locally Changes Everything

What to do and how to stay safe: Nvidia

• Review the OpenShell Reference Architecture documentation to understand how to implement fixed operational limits for AI agents in your environment.
• Evaluate whether your current agent deployments require external monitoring solutions to detect and quarantine boundary violations in real time.
• Check with your security vendors to confirm if their tools support integration with Nvidia’s OpenShell and Sentry frameworks for enhanced governance.
• Assess the compatibility of edge devices, such as RTX Spark laptops, with centralized safety policies to ensure consistent control across all deployment nodes.

Step-by-step guide: Mobile Device Security: How to Protect Data Beyond Passwords

General security guidance from the Threat Times newsroom. It is not confirmed advice from the organisations named in this story.

Frequently asked questions

What is the OpenShell Reference Architecture?

It is a design released by Nvidia for running AI agents within fixed safety limits, combining open-source runtime software with hardware-based monitoring across enterprise and edge systems.

How does Nvidia Sentry protect AI agents?

Running on BlueField-4 data processing units, Sentry monitors agent behavior from outside the main system and can quarantine or stop agents that attempt to exceed their software boundaries in milliseconds.

Which companies are working with Nvidia on this architecture?

Collaborators include Anthropic, OpenAI, Meta, Microsoft, Google Cloud, AWS, Cohere, Perplexity, Reflection and SpaceX Super Intelligence, along with various security vendors adding support.