Nvidia OpenShell architecture isolates rogue AI agents using hardware controls

Nvidia announced the OpenShell Reference Architecture at its GTC Berlin conference on Wednesday, providing a design for running AI agents within strict operational limits. The framework is available now and targets enterprise infrastructure through to edge machines. As TechSec Feed reported on 2026-10-22, Nvidia previously partnered with AWS and Canonical on this governance framework, and this release expands that foundation into a full hardware stack.

What is the technical design?

The architecture builds on OpenShell, an open-source runtime software released under the Apache 2.0 licence. It sets a boundary around agent actions. Models run on Nvidia’s Vera Rubin systems, while the software driving the agents operates on Vera CPUs. BlueField-4 data processing units execute Nvidia Sentry, which monitors agent behaviour from outside the main system. If an agent attempts to breach its software boundary, Sentry quarantines and stops it in milliseconds.

Who is adopting this framework?

Anthropic, OpenAI, Meta, Microsoft, Google Cloud and AWS are working with Nvidia on the design. Agents from Cohere, Perplexity, Reflection and SpaceX Super Intelligence can also run on the infrastructure. Security vendors including 1Password, Check Point, CrowdStrike, Cisco, Fortinet, Okta, Palo Alto Networks and Zscaler are adding support. SAP is integrating it into Joule Studio, and ServiceNow enforces actions for its Arc agent using OpenShell.

How does this affect operations?

Jensen Huang, Nvidia’s founder and chief executive, stated that AI capability and safety must advance together as enterprises deploy billions of agents. He noted that safety and control must be built into the computing architecture. Huang also said last week that he would cut off customers who build unsafe AI. The design covers edge machines including RTX Spark, DGX Spark and DGX Station systems, allowing agents to run locally on laptops.

What is the operational impact?

Nvidia claims Sentry raises agent capacity by up to 20% while enforcing boundaries. BBVA and Santander are exploring the technology to strengthen agent safety and governance. Anthropic has paired its Claude Managed Agents service with OpenShell since September, allowing companies to limit what Claude agents do and log every action. This integration provides a mechanism for auditing agent behaviour in production environments.

Questions to ask your suppliers

Organisations should verify whether their AI vendors support hardware-enforced agent boundaries and logging. Ask if the supplier’s agents can run within isolated runtimes that prevent lateral movement. Determine who inside your security team will review agent logs and define acceptable behaviour limits. Ensure your procurement process requires evidence of independent safety testing for any AI tools integrated into your core infrastructure.

Background: AI agents

AI agents are software systems that perceive their environment, make decisions, and execute actions to achieve specific goals. They move beyond generating text to performing tasks like coding, scheduling, or data analysis. You must treat them as active participants in your network, not passive tools.

Read the full guide: AI Agents Examples: How Autonomous Software Changes Security

What to do and how to stay safe: Nvidia

• Review current AI agent deployments to identify those lacking hardware-enforced boundaries or detailed logging.
• Verify that your cloud providers and AI vendors support isolation frameworks like OpenShell for agent execution.
• Define strict operational limits for AI agents, including data access and external communication permissions.
• Monitor vendor announcements for updates on agent safety features and integrate them into your risk assessments.

Step-by-step guide: Browser Security Settings for Small Business Teams

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

Frequently asked questions

When was the OpenShell Reference Architecture released?

Nvidia announced it at the GTC Berlin conference on Wednesday, and it is available immediately.

Which hardware components are used to monitor agents?

BlueField-4 data processing units run Nvidia Sentry software to watch agent behaviour from outside the main system.

Can this architecture run on edge devices?

Yes, the design covers edge machines including RTX Spark, DGX Spark and DGX Station systems.