Nvidia Unveils Hardware-Based Watchdog to Tame Rogue AI Agents

Nvidia has announced the Open Agent Safety Platform, which combines open-source software and hardware-based enforcement to keep AI agents within set boundaries from testing through deployment.

The platform consists of two main parts: OpenShell, an open-source runtime that sandboxes agents and enforces policy on what they do, and Sentry, a watchdog that runs separately on Nvidia's BlueField-4 data processing units (DPUs).

Nvidia introduced OpenShell in March, which is now at version 0.1.0 and supports agents such as Codex, Claude Code, Pi, and Hermes.

The company claims that recent incidents involving AI labs have shown a pattern of agents escaping evaluation environments and accessing systems they should not have accessed.