NVIDIA officially announced on November 10, 2026, that it is donating ‘OpenShell,’ an open-source runtime technology for managing and securing autonomous AI agents, to the Cloud Native Computing Foundation (CNCF). This move is a strategic step to expand the Kubernetes-based AI factory ecosystem and establish open infrastructure standards for running autonomous AI agents safely and efficiently across large-scale cloud and on-premises environments.
NVIDIA unveiled these plans to bolster security and control infrastructure for autonomous AI agents at its flagship conference, ‘KubeCon + CloudNativeCon.’ OpenShell, now accepted as a CNCF Sandbox project, is an open-source runtime that governs the operational permissions and execution environments of autonomous AI agents. By applying deterministic control rules when AI agents utilize external tools, access inference services, or execute workloads, it fundamentally prevents security policy violations and privilege misuse. Furthermore, it provides an extensible architecture that allows companies to easily integrate computational resources, inference providers, credential stores, and monitoring systems into their existing enterprise environments. According to NVIDIA, OpenShell has been adopted by hundreds of organizations and has recorded tens of thousands of community contributions in less than a year since its inception. Jim Zemlin, Executive Director of the Linux Foundation, stated that this donation will serve as a catalyst for community-led advancements in agent security, which is a critical infrastructure for the next generation of cloud-native applications.
Collaboration with Google’s ‘Agent Substrate’ and Kubernetes Compute Optimization
In addition to the OpenShell donation, NVIDIA is stepping up its technical contributions to the ‘Agent Substrate’ framework, a project initiated by Google and also adopted by the CNCF as a Sandbox project. Agent Substrate is an open-source platform designed to operate long-running AI agent workloads at scale within Kubernetes environments. NVIDIA has optimized this framework to interoperate with its model inference serving technology, ‘Dynamo,’ and its shared cluster resource allocation tool, ‘KAI Scheduler.’ This implementation allows AI agents to release compute resources immediately upon entering an idle state while retaining their previous work environment. NVIDIA and Google plan to jointly develop integration technology for OpenShell and Agent Substrate, enabling enterprises to ensure both agent security and execution control in heterogeneous environments spanning cloud and on-premises infrastructure.
New Components for DSX AI Factory Platform: NVKE and NVKR
New components for the NVIDIA DSX AI Factory Platform have also been launched to help companies building AI factories easily manage managed Kubernetes services while retaining control over their infrastructure and platforms. ‘NVIDIA Kubernetes Engine (NVKE)’ and ‘NVIDIA Kubernetes Recipes (NVKR),’ released in public preview, provide the open-source components necessary for building managed Kubernetes services on GPU infrastructure. These tools offer digitally signed, version-locked images and fully automate cluster lifecycle management.
Upon release, NVKE and NVKR secured support from a broad ecosystem of global cloud and infrastructure partners, with major firms committing to integration with NVIDIA technology.
- Cloud and Infrastructure Solution Companies: Mirantis, Netris, OpenNebula Systems, Rafay, Spectro Cloud, and vCluster
- Data Center and Telecommunications Infrastructure Companies: Firmus and SK Telecom
Improving Inference Startup and Recovery Performance with ‘NVIDIA Snapshot’
NVIDIA also introduced ‘NVIDIA Snapshot’ to resolve latency issues that occur when provisioning new inference capacity in Kubernetes-based AI factories. Model loading and warmup procedures are major contributors to delays in scaling large-scale inference services or agent workloads, as well as in emergency recovery. Provided as an open-source module, NVIDIA Snapshot restores checkpoint replicas in seconds without the need for model reloading or separate warmup stages, ensuring high-speed acceleration and flexible horizontal scaling.
Empirical data was also released from initial design partners Crusoe and Saturn Cloud, who validated the snapshot technology on NVIDIA HGX B300 system-based workloads. In the Crusoe Cloud environment, the time required to prepare the GLM-5.3 model for operation was reduced from approximately 8 minutes to 31 seconds. Furthermore, in benchmark tests on Saturn Cloud’s vLLM-based system, a token factory platform, initial startup times were reduced from the previous range of 2–8.5 minutes to between 6 and 34 seconds.
Technical Challenges and Future Considerations
Despite positive assessments regarding the establishment of security standardization for autonomous AI agents, the newly released OpenShell and its related tools face remaining technical challenges. Since current announcements are based on data provided by NVIDIA’s official blog, independent verification is required to determine whether practical security and stability can be maintained across diverse multi-cloud environments in the future. Moreover, as a project starting in the CNCF Sandbox stage, the key focus remains on the process of garnering broad community contributions and ensuring open-source governance alignment to successfully graduate into an incubating or graduated project.
Official Sources
This report is based on the official press release issued by NVIDIA on November 10, 2026.
- NVIDIA Official Blog: NVIDIA Expands Open Source Commitment With OpenShell Donation and New AI Infrastructure