Building secure agent workspaces with NVIDIA OpenShell Reference Architecture

• Always-on, Autonomous Agents, Now Safe for the AI Factory
  • Protecting Against
  • Outcomes
  • Maturity Model — Phase I and Phase II
• What Is Secure Agent Workspace?
• What Secure Agent Workspace Is Not
  • Workspace isolation and tenancy choices
• Where Secure Agent Workspace Fits
  • Three Timescales of Sandbox Lifecycle
• Reference Architecture
  • Typical Agent Execution Loop
  • Secure Agent Workspace — Where Everything Fits
  • Secure Agent Workspace — Seven Logical Planes
    • Architectural Invariants the Design Preserves
  • Network Architecture
  • Trusted Access Broker — What It Is and What It Must Satisfy
  • Credential Proxy — Where Tokens Live
  • Inference — A Hard Dependency, Not a Workspace Feature
• Confidential Computing as an Additive Inference Serving Profile
• Agent Blueprint Patterns
  • Posture Rules
• Enterprise Tool Access Model
• Security And Governance Model
  • Control Surface — Contracts, Implementation and Integration Boundaries
  • Baseline and Phase I Workspace Controls
  • Phase II Agent Runtime Controls
  • Threat Model — What If the In-VM OS Is Compromised?
• Operating Properties
• Deployment Tiers
  • NVIDIA Sentry Integration Profile
• Red Hat OpenShift Reference Implementation
  • Reference Shape
  • GitOps Policy Model
  • Phase I: Deployable Managed VM Controls
  • Phase II: Runtime-Enforced Agent Policy
  • Optional Linux pod profile — OpenShift sandboxed containers
  • Optional Confidential Inference Profile
  • Optional DPU networking and infrastructure visibility
  • Storage Notes
  • Operating Invariants
  • Why This Matters
• Microsoft Azure Reference Implementation
  • Reference Shape
  • Identity and Provisioning Model
  • Phase I: Deployable Managed VM Controls
  • Phase II: Runtime-Enforced Agent Policy on Azure
  • Storage and Key Management Notes
  • Operating Invariants
  • Why This Matters

Notices

• Notices
  • Notice
  • Trademarks
  • Copyright