Security & trust

Granular enough for your CISO. Invisible to everyone else.

Where DouJou runs has its own page. This one is about what it enforces once it's running — who can see what, what gets written down, and what happens when we find our own mistakes.

Who can see what

Access control that’s enforced, not requested.

DouJou runs on the identity and permissions your company already has — it doesn’t introduce a second, parallel set of rules to keep in sync.

1

Your identity, not a new one.

DouJou reads permissions from the identity system your company already runs — no separate login to provision or forget to revoke.

2

People see what they’re already allowed to see.

An answer is only built from documents the asking user already has access to — DouJou never widens someone’s reach.

3

Least privilege by default.

New data sources are connected read-only unless a task specifically requires more — access is granted deliberately, not by default.

4

Revoked in one place.

Remove someone’s access in your identity system, and their access to DouJou disappears with it — nothing left over to clean up.

Every action, logged

Built to be shown to an auditor, not taken on faith.

100%

of requests and responses logged

What was asked, what left the building, and what came back.

0

unlogged paths to an outside model

There isn’t a side door — every call to an outside AI model runs through the same checked path.

1

place to look when something needs explaining

A single, exportable trail — not logs scattered across whichever tools people happened to use.

A worked example, not a policy statement

We audited ourselves and didn’t like what we found.

The claim “your data is redacted and logged before it reaches any AI model” is only worth what the engineering behind it is worth. So here is how we actually treat it, using a real example rather than an assurance.

While reviewing an unrelated change, we audited every path in the product that sends data to an AI model. We found several that weren’t applying the governance we describe — including one inside the code path we considered our own reference implementation. No customer reported them. We found them by looking.

What we did about it

Closed every one of them in the same release and shipped it to all production deployments. Then made the guarantee structural instead of careful: an automated check now fails the build if any new code sends data to a model without going through the governance layer. It has to be explicitly and visibly exempted, with a stated reason, or the build stops.

Why that check exists

Because it immediately found several more paths that our own manual review had already walked past. The durable fix was never “be more careful” — it was making the failure impossible to introduce silently.

This is what a vendor finding its own defects looks like — the alternative isn’t a product with no defects, it’s a product where nobody went looking. If you’re evaluating us, that distinction is a fair thing to probe. We’d rather you did.

For your security reviewer

The short version, in their language.

Whoever signs off on this will ask the same questions every time. Here are the answers up front, so your champion doesn’t have to come back to us for them.

Where it runs

  • Single-tenant — one dedicated instance per customer
  • Installed inside your own AWS account
  • Connects to and inventories AWS, Azure and Google Cloud
  • No shared multi-tenant data plane
  • Your existing network and cloud controls still apply

What’s enforced

  • One enforcement point on every outbound model call
  • Unconfigured workspaces default to guarded and masked
  • Sensitive values stripped before egress, restorable only inside your boundary
  • Permissions read from your existing identity provider
  • Data sources connected read-only unless explicitly widened
  • The build fails on any ungoverned model call

What’s recorded

  • Every request and response, exportable
  • What was redacted, and what left the boundary
  • Approvals, and who granted them
  • Model, token and cost usage per request
  • No customer data used to train any model

Anything on this list we can walk through against a live deployment rather than a datasheet.

Compliance alignment

Built around the principles your compliance team already answers to.

DouJou doesn’t replace your compliance programme — it gives it something concrete to point at instead of a verbal assurance from an AI vendor.

Data residency by designLeast-privilege accessFull request/response audit trailBreach-notification-ready loggingRight-to-audit access for your team

Next step

Bring your hardest question.

The security review is the part of this we most want to have — not the part we’re trying to get past.