Security and trust
What a security review will ask, answered.
Where your data goes, what the model may do, what is recorded, who can see what, and what we hold. Written for the people who approve AI before it goes live.
Updated
Where does our data go?
Only where you agree it can. We draw the boundaries with you before anything is built, and your data never trains anyone else’s model. For work that cannot leave the building, it runs on hardware you control: nothing reaches into your network, and nothing leaves it without your say.
- Boundaries firstDrawn with you before anything is built.
- Never for trainingYour data never trains anyone else’s model.
- Your hardware, when it must beServers you own, or machines we source and build to order.
- Nothing reaches inAnd nothing leaves without your say.
The rules every system runs under
What the model may do.
The model is treated as untrusted. It can ask for a tool, and every call is checked before it runs. It reads before it writes, and anything that changes data or goes to a customer waits for a person.
- Every boundary has a name.Data crosses only at named points, each with its own check.
- The model is untrusted.It asks for tools; every call is checked before it runs.
- Read only by default.Reading, drafting and filling in forms can run on their own.
- A change needs a person’s approval.Sending, booking or spending waits for someone who sees exactly what will happen; a routine message, like a reminder, is approved once.
- Frontier models stay off.Until you switch them on: then only what a question needs, stripped of names, with no tools, and logged.
What is recorded?
Every action a system takes is on record, with its reason, so anyone can follow what it did and any change it made can be undone. Where you work under examiners or auditors, every question, every record looked up and every answer is kept where they can see it.
- Every action, with its reasonSo anyone can follow what it did.
- Changes can be undoneBecause each one is recorded as it is made.
- Questions and answers keptWhere your examiners and auditors can see them.
- A written list for each agentWhat it may do on its own, and what needs a yes.
Who can see what?
Each person finds only what they are already allowed to see. Permissions come from your own systems, so when someone loses access to a folder, the AI stops finding it for them too. Every lookup is made as the person asking, and we agree the list of sources with you before anything is read.
- Your permissions, not new onesTaken from the systems that already hold the data.
- As the person askingEvery lookup is made with their access, not a master key.
- Sources agreed firstWhich drives, mailboxes and systems, and which stay out.
- Answers show their sourceEvery answer links the document or record it came from.
What do we own?
Everything we build for you: the code, its rules, its records and its data, to keep, audit, switch off or move. If we stop working together, it stays yours.
- The codeYours to keep, read and change.
- The rules and recordsTo audit whenever you want.
- The dataWhere you keep it, under your controls.
- The off switchYou can switch it off or move it at any time.
What do you hold?
subinternal is SOC 2 compliant, and the report is available on request. HIPAA has no certification for anyone to hold, and we are not FedRAMP authorized. For that work, the system runs inside the boundary your own programs and authorizations already cover, and your reviewers approve the design there before it goes live.
What do you agree to in writing?
A data processing agreement or a business associate agreement, where the work calls for one, agreed with each client. What happens to your data when the work ends is set in the same contract. If a security incident affects your data, we tell you within 24 hours of confirming it.
- DPA and BAASigned where the work calls for one, agreed with each client.
- When the work endsWhat happens to your data is set in your contract.
- Incidents within 24 hoursYou hear from us within 24 hours of an incident being confirmed.
- Report a security issueWrite to security@subinternal.com.
Questions
Anything else your review needs, ask us and a person will answer.
Do you train AI models on our data?
No. Your data goes only where you agree it can, and it never trains anyone else's model.
Are you SOC 2 compliant?
Yes. subinternal is SOC 2 compliant, and the report is available on request.
Are you HIPAA certified or FedRAMP authorized?
HIPAA has no certification for anyone to hold, and we are not FedRAMP authorized. For that work, the system runs inside the boundary your own programs and authorizations already cover, and your reviewers approve it there before it goes live.
Will you sign a DPA or a BAA?
Yes, where the work calls for one. We agree the terms with each client.
How do we report a security issue?
Write to security@subinternal.com.
Can we use a frontier model like GPT or Claude?
Yes, if you choose to. It is off by default. When you switch it on it is sent only what a question needs, stripped of names, with no tools, and every call is logged.
Who can see our data?
Only the people already allowed to see it in your own systems. Permissions come from those systems, every lookup is made as the person asking, and we agree the list of sources with you before anything is read.
What happens to what you build if we stop working together?
It is yours: the code, its rules, its records and its data, to keep, audit, switch off or move.