How NVIDIA OpenShell Reference Architecture Secures and Governs AI Agent Workloads at Scale
Autonomous agents can work for hours on complex tasks, accessing files, tools, the internet, and spawning other agents to collaborate with. Executing complex...
First seen by Oriel on .
AI summary
NVIDIA presents OpenShell as an infrastructure-level runtime for governing untrusted AI agents, paired with Vera CPUs and optional BlueField-4 and NVIDIA Sentry hardware protections.
- OpenShell combines a lifecycle and policy Gateway, an external Supervisor for network requests and credentials, an isolating Sandbox, and a policy prover that checks permissions against organizational limits.
- Supervisor middleware and Gateway interceptors provide extension points for inspecting agent traffic and adding deployment-specific governance; identity, authorization, data protection, and auditing systems can integrate at enforcement points.
- NVIDIA says Vera’s 88 Olympus cores, coherency fabric, PCIe Gen 6, and LPDDR5X deliver up to 2x agent-workload performance versus x86, up to 3x core-to-core bandwidth, and up to 1.2 TB/s memory bandwidth.
- CoreWeave reports sandbox startup was more than 3x faster on Vera, with over 11,000 concurrent agent environments per rack; Terminal Bench results showed 1.8x faster performance across 20 agent tasks.
- For additional protection, NVIDIA Sentry on BlueField-4 monitors agent activity in an independent hardware trust domain and can quarantine agents within milliseconds when it detects threats or policy violations.
#ai agents #openshell #nvidia vera #confidential computing
Generated by gpt-5.6-sol. Check the original for details.
Aliases
| Source | External ID | URL |
|---|---|---|
| NVIDIA Developer Blog | https://developer.nvidia.com/blog/?p=123932 | https://developer.nvidia.com/blog/how-nvidia-openshell-reference-architecture-secures-and-governs-ai-agent-workloads-at-scale/ Read discussion |