Skip to navigation

NVIDIA Agent Policy Fabric Overview

View as Markdown

NVIDIA Agent Policy Fabric (APF) lets your organization decide what its AI agents may do, once, and have every sandbox follow it. Security and compliance teams write and sign policy, and APF verifies it, hands it to the agent runtime, and records which policy allowed or blocked each action.

APF is under active development and these pages document the main branch of the APF repository. Pages and sections marked In Development describe planned capabilities that are not yet in the repository. Refer to Support and Compatibility for the status of every capability.

Skip Ahead

Use Cases

  • Let a coding agent read code without letting it write. Allow repository reads and approved inference endpoints, and deny everything else. Refer to Govern a Coding Agent.
  • Keep credentials and personal data inside the sandbox. Detect secrets and personal data in outbound requests, then deny or redact them. Refer to Protect Sensitive Data.
  • Change agent policy without touching each host. Edit a document, re-sign the bundle, and reload it, and every governed runtime picks up the new revision. Refer to Update a Signed Policy.
  • Refuse policy that someone tampered with. Any change to a signed bundle causes APF to reject it. Refer to Reject a Tampered Bundle.

Refer to Use Cases for more scenarios and the controls each one uses.

Core Concepts

  • APF decides, the runtime enforces. APF verifies signed bundles and projects them into runtime-native policy. An optional decision layer evaluates each request against that policy. The agent runtime, NVIDIA OpenShell today, builds the sandbox and carries out each decision.
  • Policy is signed outside the runtime. Teams author and sign bundles offline. The services that read and enforce a bundle cannot change it.
  • Every decision is traceable. Each sandbox carries the identity of the bundle it was admitted under, and each decision is recorded with the revision of the policy that made it.

Refer to Why APF, Architecture, and The Governed Path for the full model.

Learn More