NVIDIA outlines hardware-backed controls for AI agent boundaries
NVIDIA introduced the OpenShell Reference Architecture to help enterprises constrain and monitor AI agents across data-center and local systems. The design assigns model execution to Vera Rubin systems, agent and tool management plus OpenShell sandboxing to Vera CPUs, and policy enforcement to BlueField-4 data processing units running Sentry, which can isolate an agent that exceeds its limits within milliseconds. NVIDIA said partners are building integrations and cloud support, and the architecture also extends to RTX Spark, DGX Spark, and DGX Station for local deployments.
NVIDIA says its design splits AI workloads among three hardware roles: Rubin systems host models, Vera CPUs manage agents and tools inside OpenShell’s restricted environment, and BlueField-4 DPUs run Sentry for separate policy checks. Sentry can reportedly cut off an out-of-bounds agent in milliseconds and raise agent capacity by as much as 20%.
The architecture also covers local setups through RTX Spark, DGX Spark and DGX Station, and supports Linux, Kubernetes, Windows via Microsoft Execution Containers, plus AWS, Google Cloud and Azure. Partners include ServiceNow, whose AI Passport can revoke access across software and hardware in seconds; BBVA and Santander are evaluating the framework.
The architecture could affect enterprises deploying many AI agents, especially in finance, retail, healthcare, manufacturing, robotics and telecom, by making containment and auditability part of infrastructure rather than add-on software. Security and IT teams may gain faster ways to limit runaway or compromised agents, while employees and customers could benefit from clearer operational boundaries. Its practical influence may depend on adoption, integration costs and whether independent testing confirms the claimed millisecond isolation and performance gains.