Nvidia bets on OpenClaw, but adds a security layer - how NemoClaw works

Nvidia has unveiled NemoClaw, a new security-focused software stack designed to address the significant privacy and security risks associated with the viral OpenClaw AI agent framework. Announced by CEO Jensen Huang at GTC, NemoClaw is built on OpenShell, an open-source runtime that keeps autonomous agents sandboxed and enforces organizational policy guardrails. The system allows OpenClaw agents—which uniquely leverage both Anthropic's Claude and OpenAI's ChatGPT while running locally—to operate safely by adding data privacy protections and making them more scalable. Nvidia developed OpenShell in collaboration with major security firms including CrowdStrike, Cisco, and Microsoft Security to ensure compatibility with existing enterprise cybersecurity tools.

NemoClaw can be installed with a single command, runs on any platform, and works with any coding agent, including Nvidia's own Nemotron model family. A key feature is its privacy router, which allows agents to securely access powerful frontier models in the cloud while maintaining local privacy guardrails. This architecture unites local and cloud models to teach agents how to complete tasks safely. Nvidia's broader vision, backed by its new Vera Rubin infrastructure, is to enable a generational shift toward "agents-as-a-service," where specialized AI agents can automate knowledge work for employees with enterprise-grade security, potentially replacing traditional software-as-a-service models.