Nvidia has introduced the Open Agent Safety Platform, a security framework designed to manage autonomous AI agents through engineering rather than government regulation. The platform was developed in collaboration with approximately 100 industry partners and aims to provide standardized safety boundaries for AI deployment.
The platform consists of two primary components: Nvidia OpenShell and Nvidia Sentry. OpenShell is an open-source secure runtime that utilizes kernel-level isolation to sandbox agents and enforce operator-defined policies. Sentry acts as an independent watchdog that operates on BlueField-4 DPUs, monitoring behavior at the silicon level.
Nvidia claims that Sentry can identify and quarantine rogue workflows in milliseconds if an agent attempts to exceed its defined software boundaries. While OpenShell is optimized for the Nvidia Vera CPU, it is designed to be compatible with third-party compute platforms from Arm and Intel.
Several organizations are already integrating the technology. SpaceXAI is utilizing the platform with Cursor and Grok models, while Anthropic is incorporating OpenShell and BlueField with its Claude Managed Agents. Additionally, Salesforce has integrated OpenShell with Slack for audit and permission management.
Other partners include SAP, which is embedding OpenShell into its Joule Studio, and robotics firms such as Figure, Gecko Robotics, and Skild AI. Scale AI is also adopting the technology to enhance its agentic infrastructure for enterprise and government clients.
