Repository navigation
Description
User Story
An NVIDIA AI lab partner uses OpenShell for sandboxed agent runtimes, and wants to run those sandboxes inside a CPU-based confidential VM environment. Today OpenShell's microVM isolation does not run in that environment, so the partner cannot deploy a protected agent runtime on top of OpenShell and is asking for recommended, supported configurations.
Problem Statement
OpenShell's microVM-based isolation does not run inside a CPU confidential VM, which appears to be a nested-isolation constraint of that environment. There is currently no documented or supported deployment path for OpenShell when the host itself is a confidential VM, and no guidance on which isolation backend to use or what the resulting security boundaries are. Specific confidential-VM technology, OpenShell version, kernel/hypervisor versions and exact error output: Not provided.
Impact / Why This Matters
This is a concrete runtime compatibility blocker: without a working isolation backend inside a confidential VM, OpenShell cannot be used at all in confidential-computing deployments, blocking protected agent-runtime use cases where the workload must stay inside the confidential boundary. Users in this situation have no documented fallback and no clear statement of what guarantees they would keep or lose by switching backends.
Proposed Design
- Reproduce and document the nested-isolation constraint that prevents microVM isolation from starting inside a CPU confidential VM.
- Provide a supported deployment recipe for running OpenShell inside a confidential VM, including the recommended isolation backend and required host/guest configuration.
- Evaluate and validate container-based isolation backends (Docker and Kubernetes) as alternatives for this environment, since they were raised as options but have not been validated.
- Ship runnable sample code/configuration for the recommended setup.
- Document the security boundaries of each supported configuration inside a confidential VM, making clear what isolation properties are provided by OpenShell versus by the confidential VM itself.
Acceptance Criteria
- The behavior of microVM isolation inside a CPU confidential VM is reproduced and the constraint is documented, including the observed failure output.
- At least one isolation backend is validated as working for OpenShell inside a CPU confidential VM.
- Documentation includes a step-by-step deployment recipe with the required configuration for that backend.
- Sample code/configuration for the validated setup is available in the repository.
- Security boundaries and trade-offs of each supported configuration in a confidential VM are documented.
- If microVM isolation cannot be supported in this environment, that limitation is stated explicitly in the documentation along with the recommended alternative.
Alternatives Considered
- Docker-based isolation instead of microVM isolation — discussed as an alternative but not yet validated inside a confidential VM.
- Kubernetes-based isolation — discussed as an alternative but not yet validated inside a confidential VM.
- Running OpenShell outside the confidential VM — not acceptable for the use case, since the protected agent runtime must remain within the confidential boundary.
Reported to NVIDIA by an AI lab partner.
Activity