NVIDIA NemoClaw: Securing and Managing Autonomous AI Agents
Overview
NVIDIA NemoClaw, announced at GTC 2026, is an open-source software stack. It makes autonomous AI agents—referred to as “claws”—safe, secure, and enterprise-ready by providing a management and security layer for the OpenClaw framework.
Key Insights
- Open-Source Foundation: NemoClaw is an open-source stack designed for autonomous AI agents.
- Enhanced Security & Management: It acts as a critical security and management layer for the popular OpenClaw framework, enabling secure, independent task execution for AI assistants.
- Bundled Technologies: NemoClaw integrates multiple NVIDIA technologies, providing a single-command installation experience.
- Declarative Policy Control: It introduces fine-grained, declarative policies to mitigate historical security risks associated with autonomous agents.
- Hardware Agnostic, NVIDIA Optimized: The stack runs on diverse hardware, optimized for the NVIDIA ecosystem, supporting always-on agent deployments.
Technical Details
NVIDIA NemoClaw bundles several NVIDIA technologies to enable secure and efficient operation of AI agents.
Core Components
- NVIDIA OpenShell: This security runtime creates an isolated sandbox for the AI agent. OpenShell monitors every file access and network request, preventing data leakage or unauthorized actions.
- Privacy Router: A core system that intelligently routes tasks. It determines whether a task processes using a local model for maximum privacy or offloads to a cloud model for advanced reasoning capabilities.
- Nemotron Models: Optimized AI models, such as Nemotron 3 Super 120B, are specifically designed for agentic workflows. These models excel at tool utilization and executing complex, multi-step instructions.
Addressing Agent Security
Before NemoClaw, autonomous agents posed significant security challenges due to their need for broad system permissions to function effectively. This broad access created a vulnerability for accidental or malicious data exfiltration.
NemoClaw resolves these risks by introducing declarative policies. These policies allow users to precisely define an agent’s permissions, such as: “The agent can read the ‘Projects’ folder and interact with GitHub, but it is blocked from accessing the ‘Finance’ folder or any other external website.” This mechanism prevents unauthorized access or data transfer.
Deployment Environments
NemoClaw is hardware-agnostic but delivers optimized performance across the NVIDIA ecosystem. It supports “always-on” AI assistants in various environments:
- Local PCs: GeForce RTX laptops and desktops.
- Workstations: NVIDIA RTX PRO systems.
- Data Centers: NVIDIA DGX Spark and DGX Station AI supercomputers.