Nvidia has launched the Open Agent Safety Platform, a free, open-source toolkit designed to improve the security and governance of autonomous AI agents. The platform operates across three layers, covering the agents themselves, the underlying compute resources, and the hardware.
The solution is built on Nvidia’s OpenShell software, which provides an isolated runtime environment for managing fleets of agents. It is optimized for Nvidia Vera CPUs but remains compatible with processors from Intel and Arm.
Nvidia Sentry, the second core component, utilizes the company's DOCA framework to run on BlueField-4 data processing units. It monitors agent behavior at the silicon level, allowing for real-time verification of actions and immediate shutdown if agents exceed their boundaries.
Major enterprises have already adopted the platform. SpaceXAI uses it to secure Cursor coding agents and the Grok model, while Salesforce has integrated OpenShell into Slack. SAP SE utilizes the software for runtime security within its Business AI Platform.
The platform is also being deployed by robotics firms, including Figure AI, Gecko Robotics, and Skild AI, to embed safety controls into autonomous systems. The toolkit is currently available for download via Nvidia’s developer resources and GitHub.
