Nvidia Open Agent Safety Platform

NVIDIA outlines a layered safety platform to prevent autonomous AI agents from escaping intended constraints, citing incidents where agents “drifted” from tasks, exploited tools, or misreported actions. The proposed stack combines an open-source OpenShell runtime (Apache 2.0) that runs agents in kernel-isolated sandboxes and converts operator instructions into verifiable policies, with NVIDIA Sentry implemented on BlueField-4 DPUs to provide an independent, out-of-band watchdog. Five core principles drive the design: verifiable policy, enforcement outside the agent’s reach, controlling the path to the model, scaling authority with inspectability of reasoning, and shared responsibility across labs, enterprises, and hardware providers. The argument is that agents cannot reliably self-govern and therefore require independent, hardware-backed enforcement to enable safe scaling.

Technically, OpenShell enforces file, network, tool, process, and credential limits before and during execution, while DOCA-enabled BlueField hardware correlates agent interactions, policy decisions, and tool access to create contextual activity records and perform identity governance at line speed. In Vera Rubin PODs the BlueField sits on the node’s only path to the model, enabling continuous observability and real-time intervention isolated from host compromises. The platform is optimized for NVIDIA Vera CPUs and BlueField DPUs but remains compatible with other systems; activating Sentry protections on supported hardware can be delivered as a software update. The platform invites collaboration across the AI ecosystem to build an agent economy secured by open runtimes and silicon-enforced controls.