NVIDIA has introduced the Open Agent Safety Platform, a comprehensive framework designed for the full-stack governance and control of AI agents, robotics systems, and agent hardware. The platform includes the OpenShell security software, which establishes operational boundaries for AI agents running on CPUs. This software is compatible with both open-source and closed-source models and supports third-party hardware, including products from Arm and Intel.
The company also launched NVIDIA Sentry, a system that continuously monitors AI agent behavior. Sentry operates on BlueField-4 DPUs and can isolate or halt an agent within milliseconds if it attempts to bypass security constraints. By utilizing an out-of-band trust domain, the system executes security policies at the chip level, independent of the agent or potential attackers, while integrating threat detection and data access protection.
Early adoption of the technology is already underway. Companies such as Anthropic, SpaceX, and Scale AI have begun collaborating with NVIDIA to integrate OpenShell into their respective AI models and agent systems. The platform is designed to maintain low performance overhead on the Vera platform, ensuring that security measures do not impede the speed at which AI agents complete their tasks.
