← Back
NVIDIA
NVIDIA releases OpenShell runtime for safer autonomous AI agents with policy-based isolation
· releasefeaturesecurityopen-sourceplatform · developer.nvidia.com ↗

NVIDIA Addresses Agent Security Gap with OpenShell

NVIDIA today announced NemoClaw and OpenShell, addressing a critical gap in the infrastructure needed to safely deploy autonomous, self-evolving AI agents. As AI agents evolve from responsive assistants to independent actors capable of persistent execution, continuous learning, and spawning sub-agents, traditional security models prove insufficient. OpenShell introduces out-of-process policy enforcement—moving security controls entirely outside the agent's reach.

How OpenShell Works

OpenShell sits between agents and infrastructure, governing execution, visibility, and data access. Key architectural features include:

  • Sandbox isolation designed specifically for long-running, self-evolving agents (not generic container isolation)
  • Out-of-process policy enforcement ensuring agents cannot override constraints even if compromised
  • Granular permission controls across filesystem, network, and process layers
  • Runtime verification of all actions before execution
  • Session isolation preventing privilege inheritance and cross-session contamination

Developers can deploy any agent—including OpenClaw, Anthropic's Claude Code, or OpenAI's Codex—unmodified inside OpenShell using a single command: openshell sandbox create --remote spark --from openclaw.

Why This Matters for Agent Deployment

Long-running agents present a fundamentally different threat model than stateless chatbots. They maintain persistent shell access, live credentials, accumulated context across sessions, and the ability to rewrite their own tooling. Every prompt injection becomes a potential credential leak; every third-party skill is an unreviewed binary with filesystem access. OpenShell addresses this by enforcing safety, capability, and autonomy simultaneously—a balance impossible with existing approaches.

Deployment Options

NVIDIA's Agent Toolkit provides the complete deployment stack—models, tools, evaluation, and runtimes—for building and optimizing production-ready long-running agents. OpenShell, built on Apache 2.0, integrates with NVIDIA Nemotron open-source models and supports deployment across NVIDIA RTX PCs, DGX Spark, DGX Station, and on-premise infrastructure.