Nvidia launches Open Agent Safety Platform to secure AI systems

Nvidia announced on Monday, September 28, 2026, the launch of the Nvidia Open Agent Safety Platform. The release consists of open-source software and a reference system design created to enforce full-stack governance across the software, compute hardware and robotics systems running autonomous AI agents.

The platform introduces two core technical components: Nvidia OpenShell and Nvidia Sentry. OpenShell is open-source secure runtime software that runs with minimal overhead on Nvidia Vera, a CPU built for agentic AI. It provides runtime boundaries to trace actions and enforce policies across both open and closed models, with support extendable to third-party processors from Arm and Intel. Sentry acts as an out-of-band hardware watchdog operating on Nvidia BlueField-4 DPUs, using Nvidia DOCA software to monitor agent behavior and quarantine unauthorized actions in milliseconds.

More than 100 organizations are collaborating with Nvidia on the platform. Anthropic is pairing Claude Managed Agents with OpenShell and BlueField, while SpaceXAI is applying the system to Cursor coding agents and Grok models. Other partners deploying or integrating the architecture include Scale AI, Salesforce through Slack, SAP, Microsoft, JPMorganChase, Citi, Figure, Gecko Robotics and Red Hat.