Nvidia OpenShell
Nvidia OpenShell is an open-source runtime for AI safety and security developed by Nvidia for executing autonomous AI agents in isolated sandbox environments. It controls agents' access to files, processes, networks, external services, and credentials through kernel-level isolation and declarative security policies.[1][2]
OpenShell was publicly previewed for Ubuntu at Computex in June 2026 as part of a collaboration between Nvidia and Canonical.[2] Version 0.1.0 expanded the runtime with policy verification, credential-protected service access, support for multi-tenant deployments, and CPU and GPU execution.[3]
Architecture
[edit]OpenShell consists primarily of a gateway, supervisor, and sandbox. The gateway manages sandboxes and their policies, while the supervisor controls communication between an agent and external services. Agent workloads execute inside sandboxes with operating-system controls restricting filesystem and process access.[3]
Policies are defined declaratively and can restrict network destinations and individual operations on services. OpenShell can also keep credentials outside the agent workload and provide them only to authorized requests. Its policy prover uses formal methods to analyze whether proposed permissions exceed defined security boundaries.[3]
OpenShell supports local and remote execution using Docker, Podman, MicroVMs, and Kubernetes.[4]
See also
[edit]References
[edit]- ↑ "Overview of NVIDIA OpenShell". NVIDIA OpenShell Developer Guide. Nvidia. Retrieved September 28, 2026.
- 1 2 "Securing AI agent workflows on Ubuntu with the new NVIDIA OpenShell snap". Canonical. June 1, 2026. Retrieved September 28, 2026.
- 1 2 3 Watson, Alex; Golshan, Ali (September 28, 2026). "Add Runtime Controls to AI Agents with NVIDIA OpenShell". NVIDIA Technical Blog. Nvidia. Retrieved September 28, 2026.
- ↑ "Support Matrix". NVIDIA OpenShell Developer Guide. Nvidia. Retrieved September 28, 2026.