NemoClaw: Securing AI Agents at GTC 2026
At the GPU Technology Conference 2026 in San Jose, NVIDIA CEO Jensen Huang took the stage for what many expected to be a typical hardware keynote. Instead, the first twenty minutes were devoted to something unexpected: AI agent security. Standing alongside OpenClaw founder Peter Steinberger, Huang unveiled NemoClaw — NVIDIA's open-source security stack purpose-built for autonomous AI agents.
"The age of AI agents is here," Huang told the packed SAP Center audience. "But agents without security guardrails are like self-driving cars without brakes. NemoClaw is the braking system for the agentic era."
What Is NemoClaw?
NemoClaw is the convergence of four technologies into a unified, open-source security platform:
- •OpenClaw — the open agent framework that has amassed over 300,000 GitHub stars since its inception, providing the core agent orchestration layer
- •NVIDIA OpenShell — a kernel-level security runtime that sandboxes agent execution with hardware-accelerated isolation
- •Nemotron 120B MoE — NVIDIA's mixture-of-experts model optimized for security policy evaluation and intent classification
- •Privacy Router — an intelligent routing layer that determines whether agent tasks should be processed locally or sent to cloud endpoints, based on data sensitivity
Together, these components form a defense-in-depth architecture where every agent action passes through multiple security checkpoints before execution.
The Partnership Story
The collaboration between NVIDIA and OpenClaw began in late 2025, when OpenClaw's explosive growth — from a weekend project to the fastest-growing open-source project in GitHub history — revealed critical gaps in AI agent security. Enterprise customers wanted to deploy OpenClaw agents for customer support, sales automation, and infrastructure management, but lacked confidence in the security model.
"We had over 300,000 stars and thousands of enterprise inquiries," Steinberger explained during the keynote. "But the number one blocker for production deployment was always the same question: how do we trust an autonomous agent with access to our systems?"
NVIDIA's answer was to bring its deep expertise in hardware-accelerated computing, security runtimes, and large language models to bear on the problem. The result is NemoClaw — a stack that can be deployed on anything from a single DGX Spark to a full DGX SuperPOD cluster.
Key Announcements from the Keynote
NemoClaw Blueprints
NVIDIA introduced six pre-built deployment blueprints covering the most common enterprise agent scenarios:
- 1.Customer Support Agent — handles tier-1 support tickets with full audit logging and PII redaction
- 2.Sales Operations Agent — manages CRM updates, lead scoring, and outbound communications with compliance guardrails
- 3.Security Operations Agent — triages alerts, correlates events, and executes remediation playbooks within defined permission boundaries
- 4.Infrastructure Management Agent — monitors and manages cloud resources with role-based access controls and change approval workflows
- 5.Code Review Agent — analyzes pull requests, identifies vulnerabilities, and suggests fixes with sandboxed code execution
- 6.Data Pipeline Agent — orchestrates ETL workflows with data classification and privacy-preserving transformations
DGX Spark Integration
NemoClaw ships with first-class support for NVIDIA's DGX Spark — the desktop AI supercomputer announced earlier in the keynote. A single DGX Spark can run the full NemoClaw stack including Nemotron 120B MoE, making enterprise-grade agent security accessible to individual developers and small teams.
Apache 2.0 License
The entire NemoClaw stack is released under the Apache 2.0 license, continuing OpenClaw's commitment to open source. NVIDIA emphasized that there are no proprietary dependencies — organizations can run NemoClaw entirely on their own infrastructure with no phone-home requirements.
Industry Reactions
The announcement sent ripples through the AI and security communities. Within hours of the keynote:
- •The NemoClaw GitHub repository received over 15,000 stars
- •The OpenClaw Discord server saw a 40% spike in new members
- •Multiple Fortune 500 companies announced pilot programs
- •The NemoClaw tag trended on X (formerly Twitter) worldwide
Gartner analyst Avivah Litan called it "the most significant development in AI agent security since the concept of agentic AI emerged. NVIDIA is essentially creating the security infrastructure layer that the entire industry needs."
What's Next
NemoClaw is available immediately as a public preview on GitHub. General availability is planned for Q2 2026, with enterprise support tiers available through NVIDIA AI Enterprise. The team plans to release additional blueprints, expand model support beyond Nemotron, and introduce a certification program for NemoClaw-compatible security policies.
For the OpenClaw community, this marks the beginning of a new chapter. The weekend project has evolved into an enterprise-grade platform with the backing of one of the world's most valuable technology companies. And it's all open source.
"Every line of NemoClaw is Apache 2.0," Steinberger said to close the segment. "No exceptions. The lobster spirit lives on."
