Build safer AI: Nvidia OpenShell uses hardware to quarantine rogue agents, adopted by Microsoft and Google
The chipmaker announced a reference design that uses dedicated hardware to stop autonomous software from exceeding its operational limits.
Key points
- Nvidia announced the OpenShell Reference Architecture at its GTC Berlin conference, making the design available immediately.
- The system uses BlueField-4 processors and Sentry software to quarantine agents that attempt to move outside their defined boundaries.
- Major technology companies including Anthropic, OpenAI, Meta, Microsoft, Google Cloud and AWS are working with the new architecture.
Nvidia has released a hardware and software design that keeps artificial intelligence agents inside strict operational limits. The company announced the OpenShell Reference Architecture at its GTC Berlin conference on Wednesday, according to The Next Web. This release expands on previous plans to secure autonomous software by adding dedicated hardware components to the safety framework.
In plain English
AI agents are software programs that can perform tasks automatically. These programs need to interact with other systems to be useful. However, this access creates risks if the software behaves unexpectedly or maliciously. Nvidia’s new design addresses this by creating a physical and software boundary around the agent’s operations.
The system uses specific chips to monitor the agent’s actions continuously. If the software tries to do something it is not allowed to do, the monitoring system stops it immediately. This approach ensures that safety controls are built into the computing infrastructure itself, rather than relying solely on software rules that might be bypassed.
The background
Nvidia previously introduced OpenShell as open source runtime software that sets boundaries for agent behavior. As Vector Update reported in October 2026, the company launched this framework to govern corporate AI agents. The new reference architecture builds on that foundation by integrating specific hardware components from Nvidia’s Vera Rubin systems.
The design runs agent models on Vera Rubin systems while the software driving the agents operates on Vera CPUs. BlueField-4 data processing units run Nvidia Sentry, a monitoring tool that watches agent behavior from outside the main system. This separation allows the monitoring software to act independently if the primary system is compromised or malfunctioning.
What changes now
The OpenShell Reference Architecture is available for use immediately. Nvidia stated that agents from Anthropic, Cohere, Google, Meta, OpenAI, Perplexity, Reflection and SpaceX Super Intelligence can run on this infrastructure. Security firms including 1Password, Check Point, CrowdStrike, Cisco, Fortinet, Okta, Palo Alto Networks and Zscaler are adding support for the system.
Jensen Huang, Nvidia’s founder and chief executive, emphasized that safety and control must be built into computing architecture. He noted that Anthropic has paired its Claude Managed Agents service with OpenShell since September. This partnership allows companies to limit what Claude agents can do and log every action they take, according to Anthropic.
Financial institutions including BBVA and Santander are exploring the technology to strengthen agent safety. SAP is building the architecture into its Joule Studio runtime. ServiceNow made its Arc personal agent generally available on Wednesday, with actions enforced by OpenShell. Nvidia said Sentry can raise agent capacity by up to 20% while maintaining security.
Key terms explained
System administrators should understand how hardware-enforced boundaries differ from software-only controls. In this context, a reference architecture provides a blueprint for building secure systems. Teams usually check whether their current infrastructure supports these dedicated monitoring components. They should verify that their agent workloads can run within these constrained environments without losing functionality. This approach helps ensure that safety measures are robust and difficult to bypass.
What to do and how to stay safe: Nvidia
- Review your current AI agent deployment strategy to identify potential gaps in operational boundaries.
- Check if your existing hardware supports the separation of agent execution and monitoring tasks.
- Evaluate whether your security tools can integrate with hardware-based enforcement mechanisms.
- Monitor vendor announcements for updates on compatibility with new safety architectures.
General security guidance from the Vector Update newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
When was the OpenShell Reference Architecture announced?
Nvidia announced the architecture at its GTC Berlin conference on Wednesday, according to The Next Web.
Which companies are working with Nvidia on this new design?
Anthropic, OpenAI, Meta, Microsoft, Google Cloud and AWS are among the companies working with it, Nvidia said.
How does the system stop an agent from exceeding its limits?
BlueField-4 units run Sentry software that quarantines and stops agents in milliseconds if they try to move outside their boundary.
