NVIDIA Open Agent Safety Platform: A Reference for Continuous In-Silicon Agent Monitoring

NVIDIA Targets AI Agent Safety with Hardware-Enforced Security Platform

NVIDIA has announced the Open Agent Safety Platform, a new reference architecture designed to address growing safety concerns surrounding autonomous AI agents. The announcement comes in response to recent industry reports of agents escaping their intended sandboxed environments. The platform aims to provide a foundational trust layer by combining a new open-source runtime, OpenShell, with hardware-level enforcement using the company’s BlueField-4 data processing units (DPUs), establishing a secure environment for developing and deploying agentic systems.

A Layered, In-Silicon Architecture

The platform’s architecture is built on a layered approach that separates the agent's operating environment from its security controls. At the software layer, NVIDIA OpenShell provides an Apache 2.0-licensed runtime that executes each agent in an isolated sandbox with kernel-level isolation. Operators define verifiable policies that dictate an agent’s permissions for files, networks, and tools. For a higher level of security, NVIDIA Sentry runs on BlueField-4 DPUs, providing an independent, “out-of-band” enforcement layer. In NVIDIA Vera Rubin POD systems, the DPU is positioned on the node’s only physical path to the AI model, allowing it to monitor agent behavior and enforce policies at line speed, completely isolated from the host CPU and the agent itself.

• Verifiable Policy: Operators define explicit agent access to files, networks, tools, and credentials before runtime.
• Out-of-Band Enforcement: Security controls are physically isolated on the BlueField DPU, making them inaccessible to the agent.
• Kernel-Level Isolation: OpenShell contains each agent within a secure, sandboxed environment to prevent breakout.
• Hardware-Accelerated Monitoring: The platform leverages the NVIDIA DOCA software framework to correlate agent activity and enforce policy directly in silicon.

Building a Foundation for the Agent Economy

By introducing hardware-level security, NVIDIA is positioning itself as a key infrastructure provider for what it terms the “agent economy.” The company draws parallels to the early internet, arguing that trust and safety mechanisms like browser sandboxing were not an impediment to innovation but an accelerator that enabled e-commerce and the modern web. The platform promotes a shared responsibility model among AI labs, enterprises, and hardware providers. By making OpenShell an open standard, NVIDIA is encouraging broad industry collaboration to build a common safety foundation, aiming to enable the next wave of AI applications by ensuring they can be deployed securely and at scale.

NVIDIA is leveraging its full hardware and software stack—from Vera CPUs to BlueField DPUs—to establish itself as the foundational trust and safety provider for the nascent AI agent economy. By placing security controls directly in silicon and 'out-of-band' from the agent, NVIDIA is making a strategic play to own the infrastructure layer where agentic AI will be deployed at scale, turning a critical safety problem into a hardware-accelerated moat.