Nvidia Unveils AI Agent Safety Platform with OpenShell and BlueField-4 Sentry

Nvidia has launched a two-layer platform to prevent rogue AI agents from escaping their test environments. The OpenShell runtime and BlueField-4 Sentry chip are designed to work together to enforce agent boundaries at both software and silicon levels.

The OpenShell software is an open-source runtime that executes agents inside sandboxed environments, isolating them from reaching systems outside their assigned scope. It enforces policies governing access to tools, data, and APIs, and records agent actions at the kernel level.

The Sentry hardware layer, built on Nvidia's BlueField-4 data processing units, performs out-of-band monitoring of agent behavior from separate hardware rather than from within the host system running the agent. It can detect boundary violations and quarantine an agent within milliseconds.

Prominent names such as Anthropic, IBM, and SAP are reported to be working with platform technologies, signaling broad industry interest in the interoperability standard being proposed by Nvidia.

Nvidia's VP Justin Boitano said that the platform 'could have stopped the breach' if it was being used in frontier labs for model evaluation early on, but no independent recreation has confirmed this assessment.