HeadsUpAI

NVIDIA Open-Sources OpenShell to Secure Enterprise AI Agents with Kernel-Level Isolation

· Updated

NVIDIA open-sourced OpenShell, a secure runtime that executes autonomous AI agents within isolated sandboxes. It uses kernel-level isolation (low-level security separating the agent from the host OS) to control an agent's ability to read files or call APIs. This release follows the launch of the NVIDIA Agent Toolkit.

As AI shifts from chatbots to agents that use computers, security is the primary bottleneck for enterprise adoption. Without isolated execution, agents risk leaking credentials or damaging production systems. OpenShell moves security to the infrastructure layer, mirroring OpenAI's Vercel Sandbox integration and Cloudflare's isolated environments.

You can adopt OpenShell under the Apache 2.0 license to build agents that comply with enterprise IT standards. Security policies are managed through declarative YAML files, defining what an agent can access or send. The runtime is compatible with the NVIDIA NemoClaw reference stack.

View the full update on github.com
NVIDIA AI
NVIDIA AI
@NVIDIAAI
X

We created OpenShell to make AI agents safe for enterprises. Built in open source so any company can adopt and trust it, this secure sandbox controls what agents can access, share, and send. Our CEO, Jensen, explains 👇 https://t.co/7EiIsxr0CG

60retweets368likes
View on X

Still wondering? A few quick answers below.

NVIDIA OpenShell is a secure, open-source runtime designed to execute autonomous AI agents in isolated environments. It acts as a sandbox that protects enterprise data and credentials by controlling what an agent can access, share, or send. This infrastructure-level security allows companies to deploy agents safely in production environments without risking host system exposure.

The runtime provides security through kernel-level isolation, which is a low-level protection mechanism that separates the agent's execution from the rest of the operating system. Administrators use declarative YAML files to set fine-grained policies that define the agent's permissions, ensuring it can only interact with specific files, APIs, or network resources as intended by the developer.

Yes, NVIDIA has released OpenShell as an open-source project under the Apache 2.0 license. This allows any company to adopt, audit, and trust the security framework for their own AI agent deployments. By making the runtime open source, NVIDIA enables a broader ecosystem of partners to extend the sandbox capabilities and integrate them into various enterprise stacks.

OpenShell serves as the underlying security runtime for NVIDIA NemoClaw, which is an open-source reference stack for building personal AI assistants. While NemoClaw provides the full application framework, OpenShell handles the critical task of sandboxing the agent's actions to ensure that autonomous behaviors remain within defined safety guardrails and do not compromise system integrity.

OpenShell is designed for enterprise developers and organizations building autonomous AI agents that need to interact with sensitive data or internal systems. It is particularly useful for those moving from experimental prototypes to production-grade deployments where security, privacy, and policy-based control are mandatory requirements for IT departments and security teams to approve agentic workflows.

Share this update