Running Nemotron 3 Super 120B on DGX Spark prompts a warning about injection risks—that reminder is very practical—there's no such thing as an absolutely safe sandbox; isolated testing is the baseline.

The tutorial is based on the open-source NemoClaw stack, integrating OpenShell and OpenClaw, providing a solution for long-term, secure deployment of autonomous AI assistants locally. Using the NVIDIA DGX Spark system as an example, it details the environment requirements (Ubuntu 24.04, Docker 28.x, Ollama, Telegram bot token, etc.) and deployment steps, which are expected to take 20–30 minutes, with an additional 15–30 minutes to download approximately 87GB of models. The core components include NemoClaw, OpenShell, OpenClaw, Nemotron 3 Super 120B, and NIM/Ollama inference deployment. The article notes that although there is strong isolation, no sandbox can fully defend against advanced prompt injection; testing should be conducted on isolated systems.