Demo library › Chapter 5
Enforce or monitor
Turn Enforce off for one policy and the same requests are allowed - and still recorded as failing their check. How to roll guardrails out without breaking a single pipeline on day one.
The problem
Nobody turns on a new blocking control across every pipeline on day one. The first question from a platform team is "what would this have blocked last week?" - and the second is "how do I turn it off if it breaks a release?"
Lands with: whoever will have to roll this out - platform owners, change managers, the person who gets paged
Before you start
- Policy sets ↗ open in the dashboard.
- The exfiltration test open in another tab.
Run it
Turn enforcement off
| Do | Say |
|---|---|
| Open the Data perimeter policy set. Turn Enforce off for aws-data-perimeter, and save. | Enforcement is per policy, per policy set. I'm switching the data perimeter to monitor-only. |
| Wait 30 seconds, then run the exfiltration test through Cloudflare CWS. | The proxy polls for its configuration every thirty seconds - no restart, no redeploy. And now the five writes are allowed. |
| Open Proxy decisions and expand the latest S3 request. | But look at the record: the policy was evaluated, it failed, and enforce was false. That's your report of what this rule would have blocked - from real traffic, before anything is enforced. |
Whybackground, if you are asked
A failed check with Enforce off is recorded, not ignored. That's what makes a staged rollout possible: run every new policy unenforced, read the decisions until there are no surprises, then turn Enforce on.
Turn it back on
| Do | Say |
|---|---|
| Turn Enforce back on for aws-data-perimeter, save, wait 30 seconds, run the test again. | Five of five blocked again. Turning a guardrail on or off is one switch, and it takes effect within a minute everywhere the proxy runs. |
The deployment's defaults
| Do | Say |
|---|---|
| Open Deployments ↗ - the lab's deployment. | Two settings matter here. Default mode is Allow: requests no policy cares about go through - right for guardrails. Default block makes it an allow-list. And On error action is Block: if a policy fails at runtime, the request is refused rather than waved through. |
Whybackground, if you are asked
The lab's policies, policy sets and deployment are all created by Terraform, from one manifest. A change made here will be undone by the next terraform apply - which, in production, is the point: the dashboard for experiments, code for the real thing.
What the audience should see
- The five exfiltration attempts succeeding with Enforce off.
- A decision record showing a failed check that was not enforced.
- All five blocked again a minute later.
Where the evidence is
Proxy decisions, and on the VM sudo cws-lab decisions.
Known limitations
Changes take up to one poll interval - 30 seconds, the minimum - to reach the proxy. If a run straight after saving still shows the old behaviour, run it again.
Part of: Guardrails for every pipeline · The data perimeter