Nvidia has launched the OpenShell Reference Architecture designed to monitor and contain AI agents, enhancing security and governance.
Nvidia's New Architecture for AI Safety
Nvidia has launched the OpenShell Reference Architecture at the GTC Berlin conference, designed to enhance the containment and monitoring of autonomous AI agents. This innovative framework integrates Vera CPUs, the open-source OpenShell runtime, and BlueField-4 chips operating a monitoring system called Sentry.
The architecture aims to enable AI agents to perform tasks within strictly defined parameters, ensuring they remain controlled and monitored. Nvidia CEO Jensen Huang emphasized that OpenShell allows agents to work efficiently while being governed by established limits. "Nvidia created OpenShell to give agents the freedom to do useful work within clearly defined boundaries—contained, monitored and under control," Huang stated.
OpenShell treats all agent workloads as untrusted, granting access only when necessary. Its structure includes a gateway for managing sandboxes and policies, a supervisor for outbound requests, an isolating sandbox for agents, and a policy prover for formal verification of permissions.
On the hardware side, BlueField-4 functions as a data processing unit, ensuring that Sentry can analyze traffic between an agent and the model independently, without the agent's input. This setup is crucial for maintaining security and oversight.
Nvidia previously integrated OpenShell and Sentry into the Open Agent Safety Platform, collaborating with over 100 launch partners. The need for such a framework arose after incidents where AI agents operated outside their designated environments, prompting concerns over security and governance.
Security firms like CrowdStrike, Palo Alto Networks, Zscaler, and Check Point are now incorporating support for this architecture into their security offerings. Moreover, European banks BBVA and Santander are considering its application for agent safety.
While the architecture is currently available, Sentry is expected to be fully operational by early next year, with additional resources still undergoing testing.
