NVIDIA’s NemoClaw Aims to Make OpenClaw More Secure
The rumors are confirmed: With NemoClaw, chip giant NVIDIA is jumping on the OpenClaw bandwagon, launching an open-source solution built on the OpenClaw platform that is specifically designed to close the security and privacy gaps that arise when operating autonomous AI agents. The system can be installed with a single command and is aimed at both developers and end users who want to run so-called “Claws” on a permanent and autonomous basis.
What is OpenClaw and where does the fundamental problem come from?
OpenClaw is a platform for personal AI assistants, founded by Austrian developer Peter Steinberger. The platform makes it possible to create autonomous AI agents that independently complete tasks, write code, and continuously evolve. OpenClaw has grown to become the fastest-growing open-source project in history.
The fundamental problem lies in the concept of these self-learning, always-active agents: they require extensive access to systems, data, and networks in order to be productive. It is precisely this access that creates significant security and privacy risks. Until now, there has been no infrastructure layer that grants agents the freedom of action they need while simultaneously keeping them within clearly defined boundaries.
“OpenClaw brings people closer to AI and helps create a world where everyone has their own agents. With NVIDIA and the broader ecosystem, we are building the Claws and guardrails that enable everyone to create powerful, secure AI assistants.” – Peter Steinberger, developer of OpenClaw
How does NemoClaw work technically?
NemoClaw builds on the NVIDIA Agent Toolkit to extend OpenClaw. The central component is the newly introduced runtime environment NVIDIA OpenShell, which functions as an open, isolated sandbox. It ensures that agents operate in a controlled environment and cannot access sensitive data or system resources in an uncontrolled manner.
Another element is a so-called Privacy Router, which regulates the flow of data between the local system and cloud-based AI models. This allows agents to use powerful frontier models from the cloud without crossing defined privacy boundaries.
For local processing, NemoClaw automatically evaluates the available computing resources and, if sufficient performance is available, installs NVIDIA’s own language model Nemotron directly on the device. This increases both privacy and cost efficiency, as sensitive data does not need to leave the local system, according to the company.
What security problems are specifically addressed?
- Uncontrolled data access: Without a sandbox, agents can access arbitrary system data. OpenShell limits this access through policy-based controls.
- Data transmission to cloud services: The Privacy Router governs which information may be sent to external models.
- Lack of behavioral governance: NemoClaw makes it possible to define binding security and behavioral rules for agents, within which they are permitted to evolve.
- Dependency on proprietary cloud models: By deploying local open-source models such as Nemotron, dependency on external providers is reduced.
Installation and Use
Installation is carried out via a single terminal command. An onboarding process then sets up the environment. NemoClaw is compatible with a wide range of hardware, from consumer devices to professional workstations.
Supported platforms include, among others:
- NVIDIA GeForce RTX PCs and laptops
- NVIDIA RTX PRO Workstations
- NVIDIA DGX Station
- NVIDIA DGX Spark
Assessment
NemoClaw addresses a structural problem that has emerged with the rise of autonomous AI agents: the more independently and continuously an agent operates, the greater the potential attack surfaces and privacy risks become. The approach of establishing a security layer directly beneath the agent platform, rather than retrofitting security into individual applications after the fact, is considered a technically sound path.
The project is currently in an early preview phase. The source code is available on GitHub, and NVIDIA positions NemoClaw as a contribution to the OpenClaw community, not as a proprietary replacement. Alongside Nvidia, OpenAI (which brought Steinberger on board), Meta (which acquired Moltbook), as well as Tencent, Alibaba, and Zhipu AI from China have already jumped on the OpenClaw hype (more on that here).
