NVIDIA is introducing OpenShell, a new runtime designed to secure autonomous AI agents by shifting security protocols from the application layer to the infrastructure itself. As systems gain the ability to reason and to act, including reading files, executing code, and evolving capabilities, application-layer risks are increasing, prompting the need for a more robust security approach. OpenShell establishes system-level policy enforcement, creating a sandbox for each agent to operate within defined constraints, preventing overrides or data leaks even if compromised; NVIDIA explains that this means security policies are out of reach of the agent. Part of the NVIDIA Agent Toolkit, OpenShell is currently in early preview alongside the open source reference stack, NemoClaw, and is being developed in collaboration with security partners including Cisco, CrowdStrike, and Microsoft Security.
NVIDIA OpenShell Secures Agents Through Infrastructure Policy
A shift in AI security is underway as NVIDIA introduces a new approach focused on infrastructure policy rather than solely on model or application-level defenses. The emergence of autonomous agents, capable of independent action and continuous self-improvement, presents escalating risks; systems can now not only reason but also execute tasks like accessing files, coding, and managing workflows. NVIDIA OpenShell, a component of the NVIDIA Agent Toolkit, is designed to mitigate these risks by establishing a secure-by-design runtime environment for these agents. OpenShell operates by isolating each agent within a dedicated sandbox, effectively separating operational layers and enforcing policies at the system level, a strategy the company describes as analogous to the “browser tab” model for web browsing.
This means that even if an agent is compromised, it cannot override established policies or expose sensitive data, as security is dictated by the infrastructure itself. NVIDIA states that enterprises can separate agent behavior, policy definition and policy enforcement, ensuring a unified and manageable oversight system for all autonomous operations, regardless of the host operating system. NVIDIA is actively collaborating with key security partners, Cisco, CrowdStrike, Google Cloud, Microsoft Security, and TrendAI, to standardize runtime policy management across the enterprise. The company has also released NemoClaw, an open-source reference stack that simplifies the deployment of always-on assistants using OpenShell and NVIDIA Nemotron models, providing users with customizable security guardrails and control over agent behavior; NVIDIA explains that since security needs vary, NemoClaw provides a reference example for policy-based privacy and security guardrails, allowing for tailored configurations. Both OpenShell and NemoClaw are currently in early preview, signaling NVIDIA’s commitment to open development and community collaboration.
NVIDIA NemoClaw Enables Customizable Agent Sandboxes
The increasing autonomy of artificial intelligence agents demands new approaches to system security, moving beyond traditional model and application-layer defenses. NVIDIA addresses this challenge with NemoClaw, an open source reference stack designed to simplify the deployment of secure, customizable agent sandboxes built upon the OpenShell runtime and NVIDIA Nemotron models. NemoClaw offers enthusiasts a foundational example for constructing self-evolving personal AI agents, or “claws,” with enhanced privacy and security features. This approach allows for granular control, enabling users to tailor security preferences to specific use cases, similar to adjusting settings on a mobile device. The reference stack facilitates the installation of always-on assistants with a single command, supporting deployment across diverse hardware, from NVIDIA GeForce RTX laptops to DGX Spark AI supercomputers. By integrating open source models like NVIDIA Nemotron with OpenShell, NemoClaw aims to provide a secure environment for self-improving agents operating in the cloud, on premises, or on personal computing devices.
Enterprise Collaboration Expands OpenShell Runtime Ecosystem
Security protocols for increasingly autonomous AI agents are receiving a boost through expanded enterprise collaboration around the NVIDIA OpenShell runtime. Rather than focusing on securing the AI model itself, the emphasis has shifted to establishing a robust infrastructure policy layer, a strategy gaining traction with key industry players. This unified approach aims to provide a single policy layer for defining and monitoring autonomous system operations, simplifying compliance and oversight for organizations adopting agentic workflows. The design philosophy behind OpenShell centers on isolating agents within sandboxes, effectively separating application-layer actions from underlying infrastructure policies; this ensures that even a compromised agent “cannot override policies, or leak credentials or private data,” according to NVIDIA documentation. This “browser tab” model, as described by the company, controls resources and verifies permissions before any action is initiated, offering a significant advancement over relying solely on behavioral prompts.
This is the “browser tab” model applied to agents: Sessions are isolated, resources are controlled and permissions are verified by the runtime before any action takes place.
