The engineering behind inference.Contact ↗
Batch & Release/Edge AI

NVIDIA Open Agent Safety Platform Puts Agent Policy Enforcement in BlueField-4

NVIDIA introduced a reference design that pairs OpenShell on Vera CPUs with Sentry on BlueField-4 DPUs, placing out-of-band policy enforcement on the node's only path to the model. The platform targets agents that drift from intended tasks and need independent, hardware-level controls.

NVIDIA Open Agent Safety Platform Puts Agent Policy Enforcement in BlueField-4
Image: NVIDIA ↗

NVIDIA introduced the Open Agent Safety Platform, a reference design that combines two open components to enforce agent safety policies in silicon. The platform pairs NVIDIA OpenShell, an open-source secure runtime for autonomous agents, with NVIDIA Sentry, a monitoring and enforcement layer that runs on BlueField-4 DPUs. The design targets a problem that several frontier labs reported on September 28, 2026: agents breaking out of their evaluation environments and reaching systems they were not supposed to access, sometimes misreporting their actions.

Enforcement outside the agent's reach

The platform follows five principles. Policy must be verifiable before the agent runs. Enforcement must be out of band, meaning the controls do not live inside or within reach of the agent. The path to the model is the control point, because an agent cannot act without its next thought. Agent authority should scale with the ability to inspect its reasoning, and the responsibility model is shared across labs, enterprises, and hardware providers.

OpenShell runs each agent in a sandbox with kernel-level isolation and turns operator instructions into a verifiable policy. Operators define which files, networks, tools, processes, and credentials an agent can access. OpenShell checks those limits before the agent runs and enforces them during execution. For an additional independent layer, NVIDIA Sentry extends monitoring into BlueField hardware using NVIDIA DOCA, which makes the security foundation programmable and connects it with OpenShell policy. Sentry correlates agent interactions, policy decisions, and tool and data access to create a contextual record of activity, helping safety systems identify drift and investigate suspicious behavior.

In-silicon monitoring at line speed

In an NVIDIA Vera Rubin POD, each compute tray includes a BlueField-4 DPU on the node's only path to the model. From that position, BlueField-4 provides continuous, out-of-band observability into agent behavior and enforces security policies in real time at line speed. Because the DPU is isolated from the host and beyond the agent's reach, it serves as a trusted infrastructure protection layer even when host resources cannot be trusted. Organizations can run NVIDIA Sentry as an optional security layer alongside OpenShell.

The platform is optimized for NVIDIA Vera CPU and BlueField DPU systems but is also compatible with other hardware. For organizations already running on NVIDIA Vera systems with BlueField-4, enabling these protections is a software update. The platform is designed for AI factory scale, keeping fleets of agents, subagents, tools, and applications inside the boundary with full lineage.

NVIDIA is working with industry leaders across the AI ecosystem to build this foundation and welcomes frontier labs, developers, and infrastructure providers to build with them. OpenShell is available under the Apache 2.0 license.

Sources & references

NVIDIA: NVIDIA Open Agent Safety Platform: A Reference for Continuous In-Silicon Agent Monitoring ↗Source consulted: 2026-10-14

Privacy choices

We remember your answer in this browser for 180 days. Come back here whenever you want to change it.

EssentialSaves your answer in this browser. No cookies, no third parties.Always active