Key points
- NVIDIA unveiled the OpenShell Reference Architecture to control AI agents within defined operational boundaries.
- The system combines existing OpenShell software with hardware components for model execution and policy enforcement.
- Partners are integrating the architecture with enterprise software to connect business policies with agent controls.
NVIDIA introduced the OpenShell Reference Architecture to help organizations manage AI agents operating across enterprise infrastructure and local computing systems. According to Help Net Security, the design builds on existing OpenShell security software by combining it with specific hardware for running models and enforcing rules. The goal is to allow agents to perform useful work while providing tools to monitor behavior and stop activity that crosses set limits.
By the numbers
- 20% increase in agent capacity reported when Sentry runs on BlueField-4 hardware.
- 3 main parts of the computing platform divide the work: Vera Rubin, Vera CPUs, and BlueField-4.
- 1 millisecond timeframe for Sentry to isolate and stop an agent crossing software boundaries.
What the figures mean
The architecture splits tasks across NVIDIA’s computing platform to separate model execution from governance. Vera Rubin systems run the AI models, while Vera CPUs manage agent software and tool usage within a restricted OpenShell environment. BlueField-4 data processing units run NVIDIA Sentry security software, providing a separate monitoring layer that checks behavior and enforces policies through hardware. Jensen Huang, CEO of NVIDIA, stated that safety and control must be built into the computing architecture as enterprises prepare to deploy billions of AI agents. He noted that with this architecture, agents can think on Rubin, act on Vera, and be governed on BlueField.
What should you do?
Managed service providers should review current AI agent deployment strategies to identify where boundary enforcement might be lacking. Teams usually need to verify that monitoring tools can detect out-of-bound actions in real time. Check whether existing security policies can be integrated with agent management platforms to ensure consistent governance. Organizations should also assess if their current hardware supports the separation of model execution and security enforcement layers to maintain operational control.
Operational impact
Operations teams should evaluate how agent boundaries align with current security policies. Suppliers need to confirm if their platforms support hardware-enforced monitoring. Internal stakeholders must understand the separation between model execution and governance layers. Management should review monitoring capabilities to ensure real-time detection of boundary violations. Teams usually coordinate with IT security to define acceptable agent behaviors. Check whether existing tools can integrate with new enforcement mechanisms. Ensure that all departments understand the controls governing agent actions.
What to do and how to stay safe: NVIDIA
- Review current AI agent deployment configurations to ensure they operate within defined software boundaries.
- Verify that monitoring tools can detect and report agent behavior that crosses established operational limits.
- Assess whether existing hardware supports the separation of model execution and security enforcement layers.
- Plan to integrate enterprise software policies with agent management platforms once vendors provide updates.
Step-by-step guide: Shadow AI Explained: 8 Questions Answered for Secure Operations
General security guidance from the BitShield News 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 by NVIDIA that combines OpenShell security software with hardware to control AI agents within defined boundaries across enterprise and local systems.
How does the architecture enforce security policies?
BlueField-4 data processing units run NVIDIA Sentry security software to monitor agent behavior and enforce policies through hardware, isolating agents that cross boundaries.
Which systems does the architecture support?
It supports large computing installations as well as NVIDIA RTX Spark, DGX Spark, and DGX Station systems for local agent operations in various industries.
