NVIDIA Developing OpenShell Runtime for Autonomous Agents
NVIDIA is building the OpenShell runtime, an open source, secure-by-design platform for running autonomous agents. Autonomous agents represent a significant advancement in AI, moving beyond response generation to taking action by reading files, using tools, writing code, and executing workflows. This new runtime aims to address the growing application-layer risk associated with continuously improving and evolving AI systems.
OpenShell functions by operating each agent within its own sandbox, effectively separating application-layer operations from infrastructure-layer policy enforcement. This architectural approach ensures that security policies are managed at the system level, making them inaccessible to the agent. By enforcing constraints on the agent’s operating environment, rather than relying on behavioral prompts, OpenShell prevents agents from overriding policies or leaking sensitive data, even if compromised. This allows enterprises to clearly delineate agent behavior, policy definition, and policy enforcement, establishing a unified policy layer for monitoring autonomous system operations.
NVIDIA NemoClaw Simplifies AI Assistant Deployment
NVIDIA NemoClaw is an open source reference stack designed to simplify the installation of always-on assistants known as OpenClaw. This stack includes an example configuration of OpenShell, providing enthusiasts with a ready-to-use framework for building self-evolving personal AI agents. The need for robust security and privacy is paramount, and NemoClaw offers a reference for policy-based guardrails, allowing users to customize control over their agents’ behavior and data handling, akin to adjusting app security preferences on a mobile device.
This reference stack enables self-evolving claws to operate securely across various environments, including clouds, on-premises infrastructure, and personal computing devices. Compatibility extends to NVIDIA GeForce RTX PCs and laptops, NVIDIA RTX PRO-powered workstations, and high-performance computing systems like NVIDIA DGX Station and NVIDIA DGX Spark AI supercomputers. NVIDIA NemoClaw utilizes open source models such as NVIDIA Nemotron, integrating them with OpenShell to create secure and customizable AI assistants.
Collaborative Ecosystem for Agent Security
Securing the rapidly advancing field of autonomous systems necessitates an integrated ecosystem. NVIDIA is actively collaborating with prominent security partners to align runtime policy management and enforcement for AI agents across the enterprise stack. These partners include Cisco, CrowdStrike, Google Cloud, Microsoft Security, and TrendAI. This collective effort aims to ensure that coding agents, research assistants, and complex agentic workflows can all operate under consistent runtime policies, irrespective of their host operating system, thereby simplifying compliance and operational oversight.
The “browser tab” model is applied to agents, ensuring isolated sessions, controlled resources, and verified permissions before any action is taken. OpenShell is currently in early preview, and its development is part of NVIDIA‘s commitment to building in the open with the community and partners. This initiative seeks to empower enterprises to scale self-evolving, long-running autonomous agents safely and confidently, meeting global security standards.
Tools We Use for Working with AI:
