Nvidia launches its OpenShell Reference Architecture, a framework to secure, monitor, and govern AI agents, and partners with Canonical, AWS, BBVA, and others

If you use AI agents in your business, their operation and security are now auditable and controllable via a standardized framework.

Nvidia has launched its OpenShell Reference Architecture, a framework designed to secure, monitor, and govern AI agents. This architecture builds upon Nvidia's OpenShell secure runtime and provides a blueprint for creating, operating, and policing AI agents, particularly on Nvidia hardware. The system allows AI agents to perform tasks within defined boundaries, with features for containing, monitoring, and controlling their actions. Key components include Nvidia Vera Rubin infrastructure for running models, Vera CPUs for agent harnesses, and Nvidia BlueField-4 DPUs to support Nvidia Sentry for behavior monitoring. The framework is being made available to partners capable of supporting Linux and Kubernetes infrastructure, including Canonical, SUSE, Red Hat, and cloud providers like AWS, Google Cloud, and Microsoft Azure. Several AI model builders, including Anthropic, Cohere, Google, Meta, and OpenAI, are also participating, with banks like BBVA and Santander evaluating the spec for their security needs. Nvidia states that OpenShell collects telemetry data by default for project improvement, with an opt-out option available.

The OpenShell Reference Architecture signifies a significant step towards operationalizing AI agents in enterprise environments by providing a much-needed layer of governance and security. This move addresses the inherent risks of AI agents exceeding their intended functions, as demonstrated by past incidents where models acted beyond their remit. By offering a unified framework for creating, monitoring, and controlling these agents, Nvidia is aiming to build trust and enable broader adoption of agentic AI technologies across various industries, including finance and enterprise software.

This development is particularly impactful for organizations hesitant to deploy AI agents due to security and control concerns. The partnership with major OS vendors, cloud providers, and AI model developers suggests a concerted effort to establish OpenShell as an industry standard. As AI agents become more integrated into business processes, the ability to ensure their predictable and secure behavior, as facilitated by OpenShell, will be crucial for realizing their full potential and mitigating associated risks.