Your database stops holding people. It holds tokens.
Every name, number and document you hold moves into one guarded place inside your own cloud. Your other systems keep a stand-in value instead. Getting the real thing back means asking, and the answer is no unless somebody has said otherwise. Every ask is written down, including the ones that were refused.
Runs in your cloud or data centre. Built on OpenBao. Not available as SaaS.
- full_name
- Nandini Sharma tok_4b81c07f
- n.sharma@acme.in fb18c9eaee99@acme.in
- phone
- +91 98765 11111 +91 0889711111
- pan
- AKTPP1234C NVGTB4837G
The shape survives, so the systems that validate it keep working. The value does not, so a stolen copy of this table is worth nothing.
If someone steals a copy of your database tomorrow, they get a page of meaningless strings. The real details are somewhere else, and reaching them leaves a record.
Five checks, in this order, on every request
Nothing reads personal data directly any more. Every request, from an application or from a person, crosses the same boundary and passes the same checks. They are numbered because the order is real: no code path can take them out of turn.
- 01
Who are you
Identity comes from the sign-on you already use, and roles come from what your identity provider says about a person. Never from anything the caller puts in a header, which is asserted by a test rather than left to convention.
- 02
What are you allowed to see
Set per role and per field. A role with no explicit grant receives a token and nothing more, so the starting position is refusal rather than a wide grant that somebody remembered to narrow.
- 03
Does this need approval
Sensitive reads stop here until a second, different person agrees. The vault refuses self-approval itself rather than leaving it to policy, and an unanswered request expires: silence is a refusal, not an indefinite wait.
⏳ HELD, AWAITING A SECOND PERSON ✓ APPROVED BY SOMEBODY ELSE - 04
Give back the minimum
A token, a masked value, or the real one, whichever the job actually needs. A support agent confirming they have the right person and an analyst building a report are not the same request.
- 05
Write it down
The audit row is written before the answer is returned, so a refused request leaves exactly the same trace as an approved one. That ordering is what makes the trail worth reading after an incident.
read · customers.phone ••••7632
Five levels, and the default is the lowest
A role with no explicit grant for a field receives a token. Not an error, not a blank, and not the value. Everything above that is something somebody decided to give.
| Level | What comes back | Who approves | Typical requester |
|---|---|---|---|
| 0 | A token | Nobody. This is the default. | Your CRM, your warehouse, your logs |
| 1 | Masked, ••••••7632 | Role only | Support confirming they have the right person |
| 2 | The real value, machine to machine | A standing rule agreed in advance | Payments, KYC, a mail provider |
| 3 | The real value, for a person to read | One approver, never the requester | An analyst or agent with a stated reason |
| 4 | Many records, a named list | Two approvers, one the data owner | A migration, an audit, a regulator request |
The parts an architect asks about first
-
Tokens that keep their shape
A PAN field gives back a PAN-shaped token, a bank account gives back an account-shaped one. This is tokenisation rather than format-preserving encryption: the token is an unrelated value in the right shape, reversible only by asking the vault, so there is no cipher inside it to attack. The systems that validate what they store keep working without being rewritten, which is usually what decides whether a vault project finishes.
-
Search and joins still work
A record can be looked up by its real value without anything being decrypted, because the vault matches a fingerprint rather than the value. The same value for the same field and subject always produces the same token, so joins that worked before tokenisation keep working after it. This is the objection an architect raises in the first ten minutes, and it is answered.
-
Nobody who runs the vault can read what is in it
Opening a value needs two halves held by different systems: the encryption key, which never leaves OpenBao, and a per-record derivation context, which lives only in Postgres and which OpenBao never sees. Someone holding a database dump is refused. Someone holding the OpenBao token is refused. Only the vault service, which holds both, can open a value, and every time it does that is written down.
-
Erasure is one action
Each person has their own encryption key, so destroying it makes every value of theirs unreadable at once. That includes copies in warehouses, exports and backups nobody controls, because those copies were never readable without the vault. Token rows are kept deliberately, so the audit trail still resolves after someone has been forgotten.
-
Changing a bank account needs a second person
A field can be gated on the way in as well as on the way out. The gate fires on a change rather than on first entry, which is the distinction that matters: onboarding a vendor is ordinary work, and quietly redirecting an established vendor's payments is the attack. A bulk import cannot be used to slip one past, and a test asserts exactly that.
-
Emergency access that costs something
Somebody mid-incident can take access without waiting for an approver, with a stated reason. It is time-boxed and expires on its own, every action taken under it is marked in the trail, the security team is alerted immediately, and the episode lands on a review queue with a deadline. Break-glass that nobody reviews afterwards is just a second set of permissions.
Stated plainly, so nobody demos something that does not exist
A security buyer who finds an unstated gap stops believing the stated claims too. These are the four you would otherwise find out in week three of an evaluation.
No independent penetration test yet
One has not been carried out. The brief for it is written and names the four claims to attack and the five attacker positions to take them from, including a compromised database administrator and a compromised vault operator. We commission the test against a customer's timetable and publish the result.
No hardware security module
The seal chains to a cloud key management service, proven end to end: no unseal keys at initialisation, self-unsealing across a restart, and a refusal to start at all when the service is unreachable. A PKCS#11 hardware device needs a custom build and is available as an engagement.
The second site ships snapshots, not live replication
A snapshot goes to the second site every few minutes, so a disaster loses at most those minutes, and promoting that site is a deliberate act rather than an automatic one. We say this plainly because the alternative reading, live mirroring, is what people assume. The drill that destroys the cluster and rebuilds it from the second site passes.
Three client libraries, not five
Go, Python and TypeScript, each with no third-party dependencies and no way to skip certificate verification. Java and .NET on request. Everything is a plain HTTP API underneath, so a language without a library is not blocked.
Before you ask your architect
Does tokenisation break search and joins?
No. A record can be looked up by its real value without anything being decrypted, because the vault matches a fingerprint of the value rather than the value itself. The same value for the same field and subject domain always produces the same token, so joins that worked before tokenisation keep working after it.
What is the difference between tokenisation and encryption?
An encrypted value still contains the original, mathematically, and a token does not. A token is an unrelated value of the right shape, reversible only by asking the vault, so there is no cipher inside it to attack and no key that, once leaked, turns your stored data back into personal data. Encryption protects data you still hold; tokenisation means you no longer hold it.
Can whoever runs the vault read what is inside it?
No, and this is a control rather than a promise. Opening a value needs two halves held by different systems: the encryption key, which never leaves OpenBao, and a per-record derivation context, which lives only in Postgres and which OpenBao never sees. Someone holding a database dump is refused, someone holding the OpenBao token is refused, and a regression test asserts each case.
How does a deletion request reach data we have already copied?
Every person has their own encryption key, and destroying it makes every value of theirs unreadable at once. That includes copies in warehouses, exports and backups nobody controls, because those copies were never readable without the vault to begin with. Token rows are kept deliberately, so the audit trail still resolves after someone has been forgotten.
Do we need a separate Aadhaar Data Vault?
No. An Aadhaar number is a registered field in Arbor Vault like any other, and the pattern an Aadhaar Data Vault describes is what the vault already is: a separate store holding the numbers, with every other system keeping a reference key in their place, access restricted to whoever is allowed it, and every read written down. Registering the field takes a steward about a minute. The prior question is whether you should be holding Aadhaar numbers at all, because for many private organisations the lawful route is offline verification or a virtual ID instead.
Is this available as a hosted service?
No, and that is the design rather than a roadmap item. The vault runs inside your own cloud account or data centre, which means we cannot reach your data, cannot be compelled to produce it, and cannot lose it on your behalf. It also means you operate it, which is why the runbook, the recovery objectives and the drill procedures ship with the product.
What is OpenBao, and what did you build on top of it?
OpenBao is an open-source secrets manager under the Linux Foundation, and the vault uses it for key custody and encryption as a service. Everything else is ours: tokenisation, the token map, deterministic lookup, the five-level access ladder, the approval workflow and per-subject erasure. OpenBao's open-source build ships no tokenisation engine, because that was an enterprise feature of the project it forked from, and filling that gap is what this product is.
How much does it cost?
Arbor Vault is quoted per deployment, and the figure depends on how many systems you are moving behind the vault rather than on how many records you hold. Talk to us and we will scope it. There is no published price and no self-service tier, because every deployment so far has needed an integration conversation before a number would have meant anything.
Last reviewed . General information, not legal advice. The vault is a separate product from the Sammati consent manager and from Arbor Lens, and all of them are listed here.
Bring us the two systems that worry you most.
A walkthrough runs against your own field list rather than a demo dataset, and ends with a written scope. There is no published price because the number depends on how many systems move behind the vault, not on how many records you hold.