NVIDIA launches Open Agent Safety Platform with runtime controls

NVIDIA launched the Open Agent Safety Platform on September 28, 2026, from Santa Clara, California, introducing a two-part governance system designed to enforce runtime controls on AI agents throughout their lifecycle NVIDIA Investor Relations.

The platform combines OpenShell, NVIDIA's open-source software providing runtime governance controls, with Sentry, a reference system design for independent agent monitoring NVIDIA Developer Blog. OpenShell governs what an agent can "see, do, and interact with" during execution, while Sentry functions as an out-of-band watchdog capable of quarantining and stopping an agent "in milliseconds" if it moves outside its defined software boundary Infosecurity Magazine.

Architecture and Deployment

Sentry is designed to run on NVIDIA BlueField-4 data-processing units (DPUs), enabling independent monitoring separate from the primary agent runtime Security InfoWatch. This hardware-based isolation provides a technical layer that operates independently of the agent software itself, allowing for enforcement of governance policies even if the agent itself attempts to override controls.

NVIDIA stated the platform is intended to deliver governance and control "from agent testing to deployment" across software, hardware, computing, and robotics systems that run agents NVIDIA Solutions.

Developer Availability

OpenShell and associated skills are available through NVIDIA developer resources and GitHub, positioning the platform as infrastructure for both testing environments and production deployments. The reference nature of Sentry suggests it can be adapted across different hardware configurations, though primary deployment emphasizes the BlueField-4 DPU architecture.

The release addresses a concrete gap in the agent economy: as autonomous systems move from research into enterprise production, the ability to enforce runtime boundaries at hardware speed—rather than relying solely on software-level policies—represents a shift toward hard containment. The millisecond-scale response time is material for time-sensitive environments like financial systems, robotics, and autonomous vehicles where agent misbehavior can compound rapidly.

This is the first major open-platform release coupling in-silicon monitoring (Sentry) with open-source governance controls (OpenShell), establishing a reference architecture for enterprise deployments requiring verifiable agent containment.