NVIDIA contributes open source tools to secure and scale AI agents
NVIDIA is donating new software to the Kubernetes community to help organizations manage, secure, and scale autonomous AI agents.
The short version
- NVIDIA is donating OpenShell, an open source runtime for securing and governing autonomous agents, to the CNCF.
- The company is also contributing to Agent Substrate, a Google-led framework for scaling long-running agent workloads.
- New tools like NVIDIA Kubernetes Engine and NVIDIA Snapshot help automate GPU cluster management and speed up model loading.
- NVIDIA Snapshot reduced some model startup times from over eight minutes down to just 31 seconds in testing.
Why it matters
As companies move from simple chatbots to autonomous agents that can take actions, they need standardized ways to set boundaries on what those agents can do. These tools help businesses run these agents more safely and without wasting money on idle hardware.
Securing autonomous AI agents
NVIDIA is deepening its commitment to open source by donating OpenShell to the Cloud Native Computing Foundation (CNCF). This nonprofit organization, which is part of the Linux Foundation, hosts KubeCon and manages Kubernetes, the system used to run applications across thousands of AI factories. By moving OpenShell into the community, NVIDIA wants to ensure the security of autonomous agents is built by a wide range of experts.
OpenShell serves as a runtime for securing and governing how AI agents behave. It provides deterministic controls, which means it allows organizations to set strict rules on what an agent can do when it uses tools, accesses inference services, or executes specific workloads. This is a critical step for companies that want to use AI but need to prevent it from performing unauthorized actions.
The software is designed to be extensible, allowing teams to connect their own compute resources, credential storage, and monitoring tools. Because it supports various deployment environments, organizations can apply the same governance rules whether they are running on-premises or in the cloud. Hundreds of organizations have already used OpenShell, which received tens of thousands of contributions in less than a year.
Scaling workloads with Agent Substrate
Managing the security of agents is only one part of the puzzle; companies also need to run them at a large scale. NVIDIA is contributing to Agent Substrate, a framework originally led by Google that is now part of the CNCF Sandbox. This project focuses on operating long-running agent workloads on Kubernetes while managing the underlying hardware efficiently.
One of the primary challenges with AI agents is the cost of keeping compute power running constantly. Agent Substrate helps solve this by preserving agent environments between tasks and releasing the compute resources when the agents are idle. This allows teams to scale their workloads across a shared pool of agents rather than keeping dedicated hardware running for every single session.
NVIDIA is ensuring that Agent Substrate works seamlessly with its own technologies, such as NVIDIA Dynamo for model inference and KAI Scheduler for resource allocation. Furthermore, NVIDIA and Google are collaborating on an integration of OpenShell and Agent Substrate. This partnership aims to help enterprises secure and govern their agents across both cloud and on-premises systems simultaneously.
Infrastructure for AI factories
To support the growth of AI factories, NVIDIA announced three new components of its DSX AI Factory Platform. These include NVIDIA Kubernetes Engine (NVKE) and NVIDIA Kubernetes Recipes (NVKR), which are now available in public preview. These projects provide the building blocks for delivering managed Kubernetes services specifically on GPU infrastructure, making it easier for teams to manage the lifecycle of their clusters.
These tools automate the management of GPU clusters and provide signed, version-pinned images to ensure consistency. Several companies, including Mirantis, Netris, and OpenNebula Systems, have already aligned with NVIDIA on plans to integrate these capabilities into their own offerings. This effort aims to make it easier for developers to deploy and scale the heavy infrastructure required for modern AI applications.
A third component, NVIDIA Snapshot, addresses the technical hurdle of bringing new inference capacity online quickly. Loading large models and warming up workloads can cause significant delays during scaling or recovery. Snapshot is a modular component that restores checkpointed replicas in seconds, allowing for faster startup and elasticity without needing to reload the entire model every time.
Real-world performance gains
NVIDIA has already begun testing these new tools with design partners to see how they perform on high-end hardware. For example, Crusoe tested the Snapshot tool on their cloud using NVIDIA HGX B300 systems to run the GLM-5.3 model. They found that the tool reduced the time it took for the workload to become ready from roughly eight minutes to just 31 seconds.
Another partner, Saturn Cloud, benchmarked Snapshot for scale-from-zero inference. They observed that the tool reduced vLLM startup times from a range of two to 8.5 minutes down to between six and 34 seconds. These results demonstrate how the software can help companies respond more quickly to spikes in demand without manual intervention.
By combining these open-source contributions with managed services, NVIDIA is positioning Kubernetes as the primary foundation for AI workloads. These tools allow developers to build more efficient infrastructure while maintaining control over their software and customer experience. The goal is to provide a shared foundation where the community can collaborate on the next generation of cloud-native applications.
What to watch
- How other companies will adopt OpenShell for agent governance.
- The progress of the NVIDIA and Google collaboration on agent security.
Try it
Look up the Cloud Native Computing Foundation (CNCF) Sandbox to see other open-source projects being developed for AI infrastructure.
Sources and links
- Officialcrusoe.ai/resources/blog/nvidia-snapshot-crusoe-cloud crusoe.ai
- CodeNVIDIA Snapshot github.com
- CodeNVIDIA Kubernetes Engine github.com
- Reportingblogs.nvidia.com: the article we sourced
First reported by blogs.nvidia.com. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.