NVIDIA open-sources OpenShell to govern what AI agents may do

OpenShell applies deterministic controls to agent activity: what tools an agent may use, which inference services it may call, and which workloads it may run. Teams can plug in their own compute, credential storage and monitoring. NVIDIA says hundreds of organisations already use it, and that the project has received tens of thousands of contributions in less than a year. As a CNCF Sandbox project, OpenShell will be governed by the organisations that build it into their products.

NVIDIA is also contributing to Agent Substrate, a Google-led framework for long-running agent workloads on Kubernetes, also accepted into the CNCF Sandbox. NVIDIA and Google are collaborating on integrating OpenShell with Agent Substrate.

For AI factories, NVIDIA released three new components of its DSX platform in public preview: Kubernetes Engine (NVKE), Kubernetes Recipes (NVKR) and Snapshot. Snapshot restores checkpointed GPU replicas in 3.6 to 42.7 seconds in NVIDIA's tests, versus cold starts that take minutes. Partners Crusoe and Saturn Cloud report similar speedups on their own workloads. NVKE and NVKR are not yet recommended for production.

The Hugging Face break-in and the DNS escape showed what happens when agent environments are not isolated. OpenShell is NVIDIA's answer for enterprises that want agents, but with a hard boundary on what those agents may touch.

How widely OpenShell is adopted outside NVIDIA's partner circle is not independently measured here. The "hundreds of organisations" and "tens of thousands of contributions" figures come from NVIDIA itself.

NVKE and NVKR are in public preview. APIs and behaviour may still change.