Nvidia Puts Agent Containment Into Silicon With OpenShell and a BlueField-4 Watchdog

Nvidia released the Open Agent Safety Platform, a combination of open-source software and a reference system design meant to keep autonomous agents inside declared boundaries from testing through production. The Russian-language channel AI ML Big Data summarized it as an open toolkit for running autonomous agents in an isolated environment, built on OpenShell, a runtime the company has been developing for about a year, with a second optional layer of control above it.

OpenShell is the software half. Nvidia's developer post describes three components: a gateway that manages sandbox lifecycles and policies, a sandbox that applies kernel-level controls over filesystem and process activity, and a supervisor attached to each sandbox that checks outbound requests against policy. The design compiles an operator's instructions into verifiable policy before the agent runs, then enforces which files, networks, tools, processes and credentials it may touch. The code is on GitHub under Apache 2.0 and supports Claude Code, Codex, OpenCode and GitHub Copilot CLI among others.

The second layer is the commercially interesting one. Nvidia pairs OpenShell running on Vera CPUs with Sentry on BlueField-4 data processing units (DPUs), so the monitor sits on hardware the agent's own process cannot reach and can quarantine activity within milliseconds. SecurityWeek describes it as a hardware-based watchdog, and StorageReview counts more than 100 launch partners.

Open-sourcing the runtime while putting the enforcement layer on its own DPUs is a familiar strategy. The free software defines the interface, and the part that is hard to replace runs on Nvidia silicon.