Nvidia Unveils AI Agent Security Platform with Runtime and Hardware Controls
Nvidia has announced an AI agent security platform that aims to prevent automated systems from overstepping their bounds. The Open Agent Safety Platform combines two main components: OpenShell, a runtime environment, and Sentry, a hardware-based reference design.
The system is designed to monitor agent activity and stop agents that attempt to access resources outside their permitted boundaries. This is particularly important as AI agents take on longer-running tasks and interact with sensitive services.
Nvidia's CEO Jensen Huang emphasized the importance of 'full-stack engineering' in ensuring safety and security. The company's platform documentation distinguishes between model safeguards, which influence an agent's behavior, and runtime controls, which enforce its permissions.
The Open Agent Safety Platform provides additional controls to prevent AI agents from accessing sensitive resources. This includes features such as credential handling, where actual service credentials are kept outside the agent workload and substituted into authorized requests after checking the destination and policy.