N/ANight Ash
Back to collection
Research

NVIDIA Introduces an Open Safety Stack for AI Agents

OpenShell provides a secure runtime and Sentry a BlueField-4 reference design, with NVIDIA describing broad availability rather than universal deployment.

Published
Abstract secure runtime with four controlled access lanes and a separate sentry isolating one suspicious agent packet.
Technical illustration

NVIDIA announced the Open Agent Safety Platform on September 28, combining OpenShell, a secure runtime for AI agents, with Sentry, a reference system built around BlueField-4 infrastructure. The company says OpenShell is broadly available and can place policy controls around agent access to models, tools, data and networks.

References