NVIDIA unveils Open Agent Safety Platform

NVIDIA introduced the Open Agent Safety Platform on Sept. 28, 2026 as an open software platform and reference system design to secure autonomous AI agents from initial testing through production deployment. The company positioned the release as a way to harden how AI agents are constrained and monitored during operation, focusing on preventing misbehavior and containment escapes in real time.

The platform is structured as a full-stack reference architecture that developers and organizations can adopt or extend. It is intended to provide a consistent approach to runtime policy enforcement and safety monitoring for long-running or autonomous agents across different environments.

A central component of the platform is OpenShell, described as an open-source secure runtime for agents running on central processing units. OpenShell establishes a kernel-level sandbox that governs what resources an agent can access and which operations it can perform while executing.

Beyond sandboxing, OpenShell enforces runtime policies that define acceptable agent behavior. By placing controls at the operating system and kernel layer, it aims to limit the impact of application-level vulnerabilities or misconfigurations and to provide a consistent enforcement point for safety rules.

Complementing OpenShell is Sentry, an out-of-band watchdog that runs on separate network or accelerator silicon. Sentry continuously monitors the behavior of AI agents and the systems on which they run, looking for signs that an agent is attempting to break out of its permitted boundaries.

Because Sentry operates on independent hardware, it can act even if software running on the main system is compromised. NVIDIA said Sentry can quarantine or stop agents that attempt to escape containment within milliseconds, providing a hardware-backed last line of defense if software controls fail.

NVIDIA is releasing the Open Agent Safety Platform as a reference design together with open-source code for key components. The company stated that elements of the software stack are broadly available to ecosystem partners and developers, encouraging third parties to integrate, extend, or build on the design.

By making the platform open-source, NVIDIA aims to enable organizations to inspect, adapt, and verify the controls that govern their autonomous agents. The reference architecture format is intended to support varied deployments while maintaining a common safety framework.

In discussing the launch, NVIDIA representatives linked the platform’s capabilities to a high-profile security incident disclosed in the summer involving Hugging Face. They said the Open Agent Safety Platform tools would have or could have prevented that breach if they had been in use within frontier labs during model evaluation.

This claim underscores the platform’s focus on guarding against agent behavior that bypasses existing safeguards during testing as well as in production. By combining kernel-level sandboxing with an independent hardware watchdog, the design targets scenarios where agents exploit weaknesses in application-level or infrastructure controls.