Jensen Huang revealed a system at GTC Berlin that enables companies to block AI agents from unauthorized networks and files. The OpenShell Reference Architecture is an open source framework that uses hardware-isolated security and a policy prover to mathematically verify resource access. By blocking all actions by default, the architecture ensures agents operate within secure sandboxes and can be hosted on a company’s own infrastructure to keep sensitive code and data internal.
The framework integrates with Anthropic’s Claude Managed Agents, which is currently in public beta. While Anthropic handles the orchestration of the agents, Nvidia provides the execution environment and security controls. This combined approach allows businesses to deploy agentic AI at scale using runtime controls that monitor and log every permitted or blocked action.