How it works

Four layers, one pipeline step, and a record nobody can quietly edit.

Four layers between an intention and a prevented outage

Most teams have the first two. The last two are where outages stop.

Standard

What must be true, in one sentence. Production data is encrypted at rest.

Policy

Where it applies, how, and what the exceptions are. Written in plain language, so your engineers and your AI read the same document.

Control

The named, testable check that proves the policy holds. CTL-DAT-0001, mapped to CIS 2.1.1 and SOC 2 CC6.1.

Guardrail

The automation that stops it happening at all. The pull request fails. The unencrypted database is never created.

The integration, in full

Added between your plan and your apply. Nothing else changes.

- run: |
    terraform show -json tfplan.bin > tfplan.json
    guardai audit --plan tfplan.json --emit-evidence chain.json --out report.html

- run: guardai publish --chain chain.json --repo-label acme-prod-infra
  if: always()
  env:
    GUARDAI_INGEST_TOKEN: ${{ secrets.GUARDAI_INGEST_TOKEN }}
if: always() matters. A failing audit is exactly the run worth recording, and without it your evidence only ever shows the good days.

Exactly what leaves your network

You can see this yourself before sending anything. guardai publish --dry-run prints the payload and transmits nothing.

Sent

  • Guardrail identifiers, such as GR-AWS-0001
  • Control identifiers from the published catalog
  • Severity, decision and a count of occurrences
  • A score and a band
  • Two hashes, this run and the one before it

Never sent

  • Source code, of any kind
  • Resource names and identifiers
  • File paths
  • Variable values, secrets or state
  • Account numbers and region detail

The dashboard supplies the titles and the framework references from its own catalog, which is why your dashboard is readable by an auditor and contains none of your infrastructure.

Why a screenshot proves nothing

Every evaluation is written to a chain. Each entry carries the hash of the one before it, so removing or editing an entry breaks every hash after it.

That is what an auditor is asking for when they say prove it held for the whole period. Not a dashboard showing today. A record where removal is detectable.

Your customer's reviewer verifies it themselves, offline, without trusting you and without contacting us. Evidence that only holds up while the vendor is still in business is not evidence.

What we will not do

A model can propose a patch. A deterministic engine decides, and the model never holds credentials that can write.

Suppressing a finding is rejected, always. A remediation that silences a detector fails harder than the finding it was hiding, because the cheapest way to close a finding is to stop looking for it.

Book a call What we check