Nvidia has introduced OpenShell Reference Architecture, a framework designed to help companies run AI agents with tighter security controls. The system combines software, network hardware, monitoring, and access policies to restrict what an agent can do while it works.
The announcement arrives as businesses look to deploy agents that can handle tasks independently but remain concerned about unauthorized actions, data exposure, and prompt injection attacks. Nvidia says OpenShell lets organizations set limits around credentials, identities, tools, and access to company systems. The framework also records agent activity for monitoring and audit purposes.
OpenShell is built around Nvidia infrastructure. The reference design uses Vera Rubin systems to run AI models, Vera CPUs for agent operations, and BlueField-4 data processing units to monitor and enforce security policies. Nvidia has named Canonical, Red Hat, SUSE, AWS, Google Cloud, and Microsoft Azure among the partners supporting the architecture.
A security stack for workplace agents
ServiceNow is applying the approach to Arc, an agent for routine work across email and Microsoft Teams. Arc is now generally available for Mac and is designed for Nvidia RTX Spark devices. Windows support is planned.
According to ServiceNow and Nvidia, Arc can review inboxes and meeting notes, identify action items, and build memory and skills over time. It can run locally on an employee’s device where possible, while using more powerful external models for complex tasks.
The companies say ServiceNow’s Super Intelligence Control Tower defines rules for each agent, while OpenShell enforces them during operation. Its AI Passport feature stores permissions and risk rules. If a passport is revoked, Nvidia’s BlueField hardware can stop a specific task and isolate it for review, while approved tasks continue.
Nvidia notes that OpenShell collects telemetry by default to improve the project, though users can opt out. The company also warns that third party models, tools, and services may apply separate privacy terms.
Sources
- Nvidia launches its OpenShell Reference Architecture, a framework to secure, monitor, and govern AI agents, and partners with Canonical, AWS, BBVA, and others – The Register
- ServiceNow and Nvidia launch Arc, a personal, long-running enterprise AI agent for productivity tasks across email and Teams, for RTX Spark and Nvidia OpenShell – ZDNET
Stay up to date
AI for content creation: the latest tools, tips and trends. Every two weeks in your inbox:
