Overview
NVIDIA released the NVIDIA Open Agent Safety Platform on September 27, 2026. It is an open-source security platform for AI agents. The platform includes the open-source runtime OpenShell, the hardware-layer Sentry, and DOCA-related technologies. Its design goal is to prevent agent behavior drift and unauthorized operations during runtime through kernel-level isolation, a zero-trust environment, and out-of-band monitoring, allowing agents to execute tasks within constrained boundaries while retaining observable monitoring channels. OpenShell provides versioned release notes and downloadable assets, adopts a relatively high release frequency, and has documentation hosted at docs.nvidia.com. NVIDIA explicitly states that OpenShell is currently alpha software, and its APIs and behavior may change without notice, so it is not recommended for production environments.
Key Features
- Kernel-Level Isolation: Constrains the operating boundaries of agents through kernel-level isolation mechanisms to prevent them from breaking through preset limits when performing operations.
- Zero-Trust Environment: Adopts a zero-trust environment design that verifies every agent action to reduce the risk of unauthorized operations.
- Out-of-Band Monitoring: Provides out-of-band monitoring channels to observe agent runtime status and promptly detect behavior drift.
- Open-Source Runtime OpenShell: Includes the open-source runtime OpenShell, providing versioned release notes and downloadable assets with a relatively high release frequency.
- Hardware-Layer Sentry and DOCA Technologies: Integrates hardware-layer Sentry and DOCA-related technologies to enhance security protection from the hardware level.
Use Cases
- Constrain AI agents to execute tasks within controlled boundaries and prevent behavior drift.
- Perform out-of-band monitoring and observability management of agent runtime processes.
- Verify agent operations in a zero-trust environment to reduce the risk of unauthorized operations.
- Use kernel-level isolation to provide a secure runtime sandbox for agents.
- Serve as an open-source platform for developers to study agent security protection mechanisms.
Pros
- Open-source platform, making it easy for developers to obtain and review the code.
- Adopts a multi-layer protection design including kernel-level isolation, a zero-trust environment, and out-of-band monitoring.
- Includes the OpenShell runtime, hardware-layer Sentry, and DOCA-related technologies, with a relatively complete technology stack.
- OpenShell provides versioned release notes and downloadable assets, with a relatively high release frequency.
- Documentation is hosted at docs.nvidia.com, making it easy to consult.
Pricing
The platform is open source and no specific pricing information is provided. For specific costs, please refer to the official website.
Summary
The NVIDIA Open Agent Safety Platform is an open-source agent security platform released by NVIDIA on September 27, 2026. It prevents agent behavior drift and unauthorized operations through kernel-level isolation, a zero-trust environment, and out-of-band monitoring. The platform includes the OpenShell runtime, hardware-layer Sentry, and DOCA-related technologies, with documentation located at docs.nvidia.com. NVIDIA explicitly states that OpenShell is currently alpha software, its APIs and behavior may change, and it is not recommended for production environments.
Version History
- NVIDIA releases the Open Agent Safety Platform (2026-09-27): It bundles the OpenShell open runtime, hardware-layer Sentry and DOCA technologies, using kernel-level isolation, a zero-trust environment and out-of-band monitoring to keep AI agents from drifting or exceeding their authority. OpenShell is alpha software and not recommended for production.