Skip to content

feat(policy): scoped policy distribution and immediate revocation at sandbox scale #5677

Open

Description

@NVIDIA-AI-Labs

User Story

An NVIDIA AI lab partner uses OpenShell for running large fleets of sandboxes (on the order of 100,000 concurrently) and needs policy integration that preserves fixed per-user constraints while distributing the relevant policies to, and revoking grants from, those running sandboxes.

Problem Statement

OpenShell's proposed policy control plane does not yet define a validated permission-lifecycle design. Specifically, the following are unresolved:

  • Scoped, asynchronous distribution of policies to running sandboxes (as opposed to replicating every policy to every node).
  • Approval interfaces for granting permissions.
  • Immediate revocation of previously granted permissions across running sandboxes.

Fixed user constraints must continue to hold while policies are distributed and grants are revoked at this scale.

Impact / Why This Matters

Without defined propagation and revocation semantics, the control plane cannot guarantee that a revoked grant stops being honored by already-running sandboxes, or that per-user constraints remain enforced during distribution. Policy propagation and revocation are distinct from sandbox provisioning and require separate design work; at ~100,000 sandboxes, naive approaches such as copying all policies to all nodes are not viable.

Proposed Design

  • Define explicit propagation and revocation semantics for policies and grants, including behavior for sandboxes that are already running.
  • Implement scoped distribution so that a sandbox or node receives only the policies relevant to it, rather than the full policy set.
  • Support asynchronous distribution while preserving fixed per-user constraints.
  • Provide approval interfaces for issuing permission grants.
  • Support immediate revocation of grants, with deny precedence and time-bound (expiring) grants.

Acceptance Criteria

  • Propagation and revocation semantics for policies and grants are documented, including behavior for already-running sandboxes.
  • A prototype demonstrates scoped policy distribution without copying every policy to every node.
  • Fixed per-user constraints are preserved across asynchronous policy distribution.
  • Approval interfaces for permission grants are available.
  • Revocation of a grant takes effect immediately on running sandboxes.
  • Deny precedence is tested and holds under failure conditions and at scale (~100,000 running sandboxes).
  • Time-bound grants are tested and expire correctly under failure conditions and at scale.

Alternatives Considered

Not provided.

Reported to NVIDIA by an AI lab partner.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    state:triage-neededOpened without agent diagnostics and needs triage

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions