Guided defense lab

Least Privilege

A guided exercise on judging whether an agent's authority envelope is minimized for its task class: effective and transitive reach, unique identity versus shared credentials, bounded JIT elevation and credential lifetime versus privilege lifetime.

Conceptual simulationAdvanced15 minutes2026-09-27
Conceptual simulation

This exercise uses authored, deterministic examples with fictional identities, resources and .example.test names. It calls no tools, APIs, agents or models, contains no credential material and does not predict how any real model or system behaves.

3/6

Learning objectives

  • Distinguish standing authority from task-scoped authority.
  • Prefer a unique workload identity over shared credentials for attribution and revocation.
  • Review the effective permission union, including inherited, group, downstream and integration reach, not one role line.
  • Delegate only the task-required subset of a user's authority, not the whole user envelope.
  • Recognize bounded, recorded, auto-expiring JIT elevation as compatible with least privilege.
  • Separate credential lifetime from privilege lifetime.
  • Keep least privilege distinct from Agent Authorization, Information Flow Control, Egress and Sandboxing.
  • Explain why an accepted envelope authorizes no concrete action and why model statements never change provisioned authority.

Prerequisites

  • Least Privilege defense article (recommended)
  • Agent Authorization lab (recommended)
Choose a scenario

Each fictional scenario shows task and model text and, separately, the application-owned facts about the task baseline, identity, effective grants, tools, integrations and lifetime. You judge whether the authority envelope itself is minimized - not whether a particular action may execute.

Authority path

The application defines what a task class needs, provisions an identity and an effective envelope, and that envelope sets the reach of tools, data, integrations and network. User task content stays variable. Every concrete action is still decided separately at execution time.

  1. Task class / required capabilitiesApplication-owned->
  2. User task contentVariable content->
  3. Workload or delegated identityApplication-owned->
  4. Effective authority envelopeApplication-owned->
  5. Tool / data / integration / network reachApplication-owned->
  6. Execution-time authorization (separate)Application-owned

Task and model text

What the user asks and what a model says about its needs or intent. It cannot create, widen or extend a grant.

  1. User task

    Summarize this week's status notes for project-a.

  2. Model note

    I only need to read the status folder.

Application-owned authority facts

Task-class baseline, identity, direct and effective grants, exposed tools and integrations, and grant or elevation lifetime as recorded by the application.

Which application-owned authority facts are authoritative for evaluating this envelope?