NVIDIA contributes OpenShell agent-security runtime to CNCF, adds three Kubernetes tools to DSX

NVIDIA announced on 10 November 2026, at the KubeCon + CloudNativeCon conference, that it is handing OpenShell, its open source software for securing and governing AI agents, to the Cloud Native Computing Foundation (CNCF). The company also said it is adding three open source components to its DSX AI Factory Platform and taking part in Agent Substrate, a Google-led project already admitted to the CNCF Sandbox.

OpenShell and Agent Substrate

CNCF is a nonprofit within the Linux Foundation. NVIDIA says OpenShell will enter the CNCF Sandbox and will be governed by those organizations that build it into their products. According to NVIDIA, the software applies fixed, predictable controls to what an agent does, letting teams enforce policy as agents call tools, reach inference services or run workloads, and connect compute, inference, credential-storage and monitoring systems of their choosing. NVIDIA says hundreds of organizations have used it and that contributions now number in the tens of thousands after less than a year. Linux Foundation chief executive Jim Zemlin is quoted describing secure environments for agents as a new infrastructure challenge for the cloud-native community.

NVIDIA describes Agent Substrate as a hardened runtime for sandboxed agents. It shares a small pool of Kubernetes worker pods among many agent sessions, saving idle sessions with their memory and files and restoring them on whichever worker is free. NVIDIA lists its upstream contributions as request queuing in the router while every worker is occupied, per-container CPU and memory ceilings in microVM sandboxes, and a vendor-independent design for pausing and resuming GPU state through Kubernetes Dynamic Resource Allocation. NVIDIA is also making the project work with its Dynamo and KAI Scheduler software and, with Google, integrating it with OpenShell.

Three components for DSX

Two of the components, NVIDIA Kubernetes Engine (NVKE) and NVIDIA Kubernetes Recipes (NVKR), are for operators offering managed Kubernetes on GPU hardware. NVKR builds node and container images from pinned source versions; NVKE uses them to create and manage clusters, taking requests through a REST API. Both are Apache 2.0 public previews that NVIDIA says may change and are not yet advised for production use. Eight companies have agreed integration plans with NVIDIA: Firmus, SK Telecom, Mirantis, Netris, OpenNebula Systems, Rafay, Spectro Cloud and vCluster.

The third, Snapshot, saves a fully started GPU workload, CPU and GPU memory alike, and lets a new pod resume from it, skipping weight loading, compilation and warm-up. NVIDIA reports that in its tests on B200 GPUs, replicas came back in between 3.6 and 42.7 seconds, 8 to 36 times quicker than starting cold.

Two design partners tried Snapshot on NVIDIA HGX B300 systems. Crusoe says GLM-5.3 became ready in 31 seconds rather than roughly eight minutes, and that keeping the checkpoint on a local NVMe drive instead of NFS brought pod readiness down to 31 seconds from 50. NVIDIA says Saturn Cloud, which it describes as a token factory platform, ran a scale-from-zero benchmark that shortened vLLM startup from between 2 and 8.5 minutes to between 6 and 34 seconds.