Resources · 67
AI agents: authorize an action and verify its effect
Bind human approval to a specific operation, limit its lifetime and account for effects already executed.
· 4 min
What this guide helps achieve
- Define the exact action
- Bind approval to parameters
- Prevent reuse
- Verify effects and recovery
- Bound each delegation
- Prove revocation and stopping
Quick check
- Do executed parameters match approved ones?
- Are expired or replayed approvals rejected?
- Does stopping leave an external effect to reconcile?
Step-by-step method
- 01
Define the exact action
Describe the object, recipient, amount or content and external effects. Classify operations by risk and reversibility. Approval to draft a message does not include permission to send it.
Deliverable: operation register and limits.
- 02
Bind approval to parameters
Show the person the decisive parameters and approval scope. Verify identity and permissions in the service executing the operation. Material changes should require a new decision.
Deliverable: per-action authorization contract.
- 03
Prevent reuse
Set an expiration, operation identifier and repeat protection. Test an old approval, a changed target and concurrent execution in a test environment. Model-generated text must not grant itself a permission.
Deliverable: negative test results.
- 04
Verify effects and recovery
Reconcile the external service result with the displayed state. Check missing responses before retrying. Record completed effects and possible compensations: stopping an agent does not unsend a delivered email.
Deliverable: timeline and recovery procedure.
- 05
Bound each delegation
Tie every worker to an identity and verifiable scope: tool, resource, operation and budget. A subtask must not automatically inherit all parent privileges. Check parameters immediately before the external effect and reject invalidated approval or policy versions.
Delegation matrix and execution-point decisions.
- 06
Prove revocation and stopping
Test revocation while an action is queued or running. Separate stopping new operations, invalidating rights and effects already produced. Retain operation ID, decision, policy version and external confirmation without copying secrets or sensitive prompts. Suspend affected actions when required evidence is missing.
Revocation timeline and confirmed effect state.
Acceptance case to reproduce
Fictional example: these inputs describe no customer or observed result.
View case inputs
{
"approved": {
"target": "A",
"amount": 100
},
"execution": {
"target": "B",
"amount": 100
}
}Expected decision
The target changed: invalidate the earlier approval and authorize the exact operation before execution. In the test, keep authorised target A beside prepared target B. Check significant parameters at execution time, then obtain approval for the exact version if it changes. A valid sign-in and an unchanged amount do not transfer approval to another recipient.
Verification procedure
- Record approved target, amount and expiry.
- Change the target in an authorized test.
- Keep the rejection and parameter-bound replacement approval.
- Have two fictional workers request the same sensitive operation and observe the shared execution-point control.
- Revoke a right in an authorized test before the next action; check rejection, user message and absence of a new effect.
- Reconcile decision, policy version and external state, then check logs contain no secrets.
Reusable worksheet
Complete with your authorised observations. These fields are a working template, not observed results.
| Field | Information to record |
|---|---|
| Operation | Object, approved parameters and external effects |
| Approval | Identity, scope, expiration and identifier |
| Recovery | External state, compensation and owner |
AI agents: control delegation, revocation and evidence
Documentary update, 12 October 2026: OWASP GenAI announced its LLM Top 10 2026 and Agent Control Standard in September. An 8 October post collects third-quarter disclosures; it describes itself as non-exhaustive and not an official OWASP publication. These references help select acceptance scenarios, but do not validate an agent or establish attack frequency.
Worked example
Illustrative situation
Fictional example: an agent proposes a discount for a test account, then changes the amount after approval.
Decision and expected evidence
The service rejects the outdated authorization. The team also tests repeats and checks interruption effects separately.
Distinguish the mechanisms
| Mechanism | Purpose | Check or limitation |
|---|---|---|
| Proposal review | Review a plan or draft | Does not establish permission for an external effect |
| Operation authorization | Approve specified parameters | Expires and does not cover a changed operation |
Management indicators
| Indicator | What it measures | First action |
|---|---|---|
| Expired approvals rejected | Reuse test results | Investigate every unexpected acceptance |
| Reconciled effects | Operations linked to confirmed external state | Resolve uncertainty before retrying |
| Checked delegations | Reviewed subtasks with identity, scope and retained decision / reviewed subtasks | Separate absent rights from missing evidence |
| Actions after revocation | New effects accepted after confirmed invalidation | Separate already-started operations from revocation failures |
Common pitfalls
- Approve without displaying the target
- Confuse stopping a process with reversing effects
Frequently asked questions
Does Stop undo every action?
It can interrupt further actions. Completed effects require verification and, where possible, a separate compensation.
Is general permission sufficient?
The service must still check identity, object and operation. Sensitive actions may require approval bound to parameters.
What happens after a timeout?
Treat the outcome as uncertain and check external state before retrying to avoid duplicate effects.
Does parent permission cover every subtask?
No. Define each actor’s rights and validate delegation at the service performing the action. An agent instruction does not create authorization.
Does the incident roundup measure my system’s risk?
No. Its scope and disclosures do not provide an attack rate. Use it to select scenarios, then gather evidence on your own configuration.
Official references
References consulted: . The method and worksheet propose checks to adapt to your context; they do not constitute certification.
