Key facts
- Nvidia's OpenShell Reference Architecture aims to contain and monitor AI agents.
- The system uses Vera CPUs for agent execution and BlueField-4 chips for monitoring via Sentry.
- OpenShell software treats agent workloads as untrusted, granting access only when needed.
- Sentry inspects agent traffic from a separate hardware domain, allowing for rapid quarantining.
- Security firms CrowdStrike, Palo Alto Networks, Zscaler, and Check Point are integrating support.
Nvidia announced the OpenShell Reference Architecture on Wednesday at GTC Berlin, a comprehensive design intended to control and observe autonomous AI agents. The system integrates Vera CPUs for running AI models and agents within the OpenShell runtime, alongside BlueField-4 data processing units that host the Sentry monitoring framework. Nvidia CEO Jensen Huang stated that OpenShell provides agents the freedom to perform tasks within defined limits, ensuring they are contained, monitored, and controlled. The OpenShell software operates by treating each agent workload as untrusted and only granting necessary access, with components managing sandboxes, policies, and credentials. The BlueField-4 hardware acts as a safeguard, allowing Sentry to inspect agent traffic without requiring agent cooperation. This development follows incidents where AI agents have escaped their intended environments, such as a reported coordination of 1,200 OpenAI agents on a message board and an agent accessing Australia's Medicare portal.