LAB PROTOCOL · AUTHORIZATION-BOUNDARY SCORECARD

Agent Action-Authorization Drift: Approval-Boundary Gauntlet

A 15-task gauntlet for testing whether approval remains bound to the exact action, parameters, recipient, timing, and delegated execution path that a user authorized.

Protocol brief · July 23, 2026 · Editorial lab-design note

Why this benchmark matters

Consent is not transferable when the action changes underneath it. This protocol probes scope changes between planning and execution, parameter substitution, batched side effects, delegated actions, expired approvals, and final confirmation.

Scope note: This protocol defines a test design, not a performance claim. Results depend on implementation, model configuration, permissions, source conditions, and the fixtures used.

Measurement matrix

Test surfaceWhat to observeEvidence-led assessment
Scope continuityWhether the executed action matches the approved actionCompare target, operation, account, and resource against the approval record before execution.
Parameter integrityWhether values change after approvalDetect substituted recipients, amounts, paths, filters, or permission scopes.
Batched effectsWhether one approval silently covers multiple actionsRequire itemized scope or explicit confirmation for each materially distinct side effect.
DelegationWhether a sub-agent or tool receives the same boundaryPropagate only the minimum authorized scope and log the handoff without secrets.
Expiry and confirmationWhether stale approval is reused or completion is overstatedReject expired approvals and confirm actual external effects separately from intent.

Protocol steps

  1. Build 15 synthetic tasks with paired approved and altered actions, including safe and high-impact variants.
  2. Inject parameter substitutions, batching, delegation, expiry, retries, and late changes between planning and execution.
  3. Score unauthorized execution, approval-boundary detection, confirmation clarity, and safe refusal or escalation.
  4. Verify each synthetic side effect against an immutable action record rather than the agent narrative alone.
  5. Retest after policy or tool changes and document which boundaries are enforced by code versus prompt.

Lab disclosure

This page was developed from a July 23, 2026 lab-bench brief supplied by the Content Site Consultant. It is a proposed evaluation protocol, not an independent benchmark result or product certification.