Each cloud gives you a way to test a guardrail first, and each has a documented gap you must cover by hand
Facts read 10 Nov 2026. Page reviewed 16 Nov 2026.
The IAM policy simulator evaluates SCPs and permissions boundaries but not RCPs; Azure Policy offers DoNotEnforce; Google Cloud offers dry-run mode and Policy Simulator for organization policy. This guide sets out what each tester covers, the documented gaps, and the order to run them in.
The IAM policy simulator covers SCPs and boundaries, not RCPs
The IAM policy simulator evaluates identity-based policies, permissions boundaries, SCPs and resource-based policies that you provide, without sending a real request. Each simulation returns allowed or denied per action and resource. You can simulate one permissions boundary at a time, and SCP testing is available when the account belongs to an organization. Cross-account simulations return per-policy-type decisions. Custom mode lets you paste a draft policy without saving it to the account.
AWS states that the simulator does not support resource control policies. A new RCP therefore has no simulator path, and the documented alternative is to attach it to a test account and review CloudTrail for Access Denied errors. The simulator also tells you whether an SCP allows or denies an action but, for security reasons, does not surface the matched statements.
Simulator results can differ from the live environment
AWS warns that results can differ from live behavior and recommends checking your policies against the live environment afterward. Documented differences include VPC endpoint policies, role chaining and multiple resource-based policies on one resource. The simulator uses no real context values from production. Treat a simulator pass as evidence of syntax and intent, and the live check as the real test.
It fills in principal and organization context keys itself, such as aws:PrincipalAccount and aws:PrincipalOrgID, so SCPs with those conditions evaluate correctly. You supply the rest. In the console, a condition key referenced only by an SCP cannot be set, so use the CLI or API for a Region-deny SCP that depends on aws:RequestedRegion. Supply a value for each key your policies reference.
Azure DoNotEnforce tests the assignment on real resources
Azure Policy takes a different route: assign the real policy in DoNotEnforce mode. The effect is not enforced on create or update and no activity log entry is written for it, but evaluation still produces compliance results, so you see which existing resources the policy would have flagged. Nothing in the mode changes the resources themselves. Do not rely on that alone.
Microsoft documents DoNotEnforce as the way to evaluate an assignment before enforcement. Remediation tasks for deployIfNotExists can still run in this mode, so check whether the assignment includes one. Read compliance results for a full evaluation cycle before switching to the default mode, because the standard compliance evaluation cycle runs once every 24 hours. Moving from DoNotEnforce to Default is a one-property change to the assignment.
Google offers two tests for organization policy and one for perimeters
Policy Simulator for Organization Policy previews a new custom or managed constraint and lists the resources that would violate it, before the policy is enforced. Dry-run mode then enforces a policy as audit only: violations are logged and the violating actions are not denied. Dry-run needs billing enabled on the project. Run the simulator on the proposed constraint and read the list of violating resources.
Dry-run supports custom constraints, managed constraints and certain legacy constraints, so some older constraints cannot be trialled this way. VPC Service Controls has its own dry run mode for perimeters, where violating requests are logged and not denied. The deny-policy and principal access boundary testing pages were not available when read on 2026-11-10, so those need a lower-environment trial. Record which tests were not possible.
Run them in a fixed order
First validate syntax with IAM Access Analyzer. Then simulate, or run the cloud native preview. Then enforce in report-only mode where one exists, and read the logs for one full cycle. Then apply to one account, subscription or project, and only then to its parent. Each step narrows the blast radius if the test missed something. Skip none of them for a policy that touches production data.
Write the expected result before each run. A simulation that agrees with a prediction is evidence, and a result you did not predict is the finding. The rollout guide places these tests inside a staged rollout. Keep the written predictions with the change record, so the approver can see what was expected, what was observed and what was left untested in a lower environment.
What each tester can and cannot show
The table below sets the three clouds side by side. Read the last column as the list of things you still have to check by another route before you widen a guardrail to production. Fill in the last column for your own estate, because an untested path is where the next incident starts, and write the manual check next to each gap before you ask for approval.
| Cloud | Preview tool | Report-only mode | Documented gap |
|---|---|---|---|
| AWS | IAM policy simulator, for SCPs and boundaries | None found on 2026-11-10 for SCPs | No RCP support; SCP matched statements not shown |
| Azure | Compliance results from DoNotEnforce | DoNotEnforce, Enroll | Remediation tasks can still run |
| Google Cloud | Policy Simulator for Organization Policy | Dry-run mode, perimeter dry run | Only some constraint types support dry-run |
Related pages
Sources
- IAM: Policy testing with the IAM policy simulator, read 10 Nov 2026
- AWS Organizations: Resource control policies, read 10 Nov 2026
- Azure Policy: Assignment structure, read 10 Nov 2026
- Azure Policy: Overview, read 10 Nov 2026
- Google Cloud: Test organization policies in dry-run mode, read 10 Nov 2026
- Google Cloud: Policy Simulator for Organization Policy, read 10 Nov 2026
- Google Cloud: Dry run mode for service perimeters, read 10 Nov 2026
Last reviewed 16 Nov 2026.