Nvidia Donates OpenShell Agent-Security Runtime to CNCF, Ships AI-Factory Kubernetes Tools

Nvidia has donated OpenShell, its open source runtime for securing autonomous AI agents, to the Cloud Native Computing Foundation, and says hundreds of organizations already run it. The company announced the move on Tuesday at KubeCon + CloudNativeCon North America in Salt Lake City. OpenShell enters the CNCF Sandbox, the Linux Foundation body that also hosts Kubernetes, and will be governed by the organizations building it into their products.

OpenShell applies deterministic controls to agent activity. Teams can enforce policies as agents call tools, reach inference services and run workloads. The runtime plugs into an organization's own compute, inference providers, credential storage and monitoring. Nvidia says the project has drawn tens of thousands of contributions in less than a year since launch.

"The cloud-native community has a new infrastructure challenge: giving agents secure environments in which to operate," said Jim Zemlin, CEO of the Linux Foundation.

What changes for agents on Kubernetes?

Nvidia is also contributing to Agent Substrate, a Google-originated framework accepted into the CNCF Sandbox for running long-lived agent workloads on Kubernetes, and making it work with Nvidia's Dynamo inference software and KAI Scheduler.

The framework addresses a structural cost problem. Agents spend most of their time idle, waiting on a prompt, a tool or a model. Agent Substrate checkpoints idle sessions — memory and filesystem included — and restores them onto whichever worker pod is free, rather than keeping one pod running per agent. Nvidia's developer blog claims 10x higher density than standard containers. Nvidia and Google are integrating OpenShell with Agent Substrate so companies can govern agents across clouds and on-premises systems.

What is Nvidia adding for AI factories?

Three new open source components extend Nvidia's DSX AI Factory Platform:

• NVIDIA Kubernetes Engine (NVKE) runs a managed Kubernetes service on an operator's own GPU infrastructure.
• NVIDIA Kubernetes Recipes (NVKR) builds the node images, containers and packages those clusters need from validated combinations of software versions and GPU generations.
• NVIDIA Snapshot restores checkpointed inference replicas in seconds without reloading or warming up the model.

NVKE and NVKR are in public preview under the Apache 2.0 license. Nvidia says they are not yet recommended for production. Mirantis, Netris, OpenNebula Systems, Rafay, Spectro Cloud, vCluster, Firmus and SK Telecom plan to build them into their offerings.

How fast is Snapshot in practice?

Design-partner tests on HGX B300 systems produced the sharpest numbers of the announcement. Crusoe cut the time for a GLM-5.3 workload to become ready from roughly eight minutes to 31 seconds — a drop of about 94 percent. Saturn Cloud cut vLLM startup times from between two and 8.5 minutes to between six and 34 seconds.

The releases follow Nvidia's opening of its Cosmos 3 models last week, continuing a pattern of donating software to open foundations while keeping its hardware and platform stack central to deployment. If Sandbox adoption proceeds, OpenShell's governance will pass to the vendors embedding it, and NVKE's move from public preview to production readiness will determine whether managed Kubernetes on Nvidia GPUs becomes a standard build for AI factories.