ARBOR LENS

You cannot protect what you never knew you had.

Lens connects to a database with read-only credentials, works out its structure by itself, samples it, and hands back a list of where the personal data is. It never copies a record out. The findings are the only thing that leaves, and the samples inside them are masked before they are written.

  • Runs in your perimeter
  • Read-only credentials
  • No outbound calls

Ships as a container or onto a throwaway host, and runs entirely inside your perimeter. When a scan finishes it reminds you that the read-only credentials can now be revoked, because that is good hygiene and we would rather model it than mention it.

In plain terms

Most organisations writing their first privacy notice are guessing at what they hold. Lens replaces the guess with a list, produced from the systems themselves rather than from a workshop.

What happens

Five steps, and you only perform the first one

Numbered because the order is real: scope is enforced before anything is read, and the findings are written last. Discovery itself is the job of the tool, so you never have to say where the personal data is.

  1. 01 · POINT

    You point it at one source

    A connection and a named database, nothing more. Read-only credentials are enough, and a dry run checks reachability and what the connector can do without reading a single record.

  2. 02 · DISCOVER

    It works out the structure itself

    Lens enumerates the collections in scope, reads a declared schema where one exists, and infers fields from samples where none does. A store that cannot enumerate scans exactly what you named rather than failing, and whichever of those happened is recorded in the findings.

  3. 03 · SCOPE

    It scans only what you named

    A credential that can reach forty databases still scans the one you pointed at. The scope rule lives in the connector base class rather than in an operator remembering to be careful, and collections, field paths and key patterns can each be excluded.

  4. 04 · SCORE

    It scores every field on three signals

    The field name, the shape of the values across the whole sample, and a published checksum where one exists. Never one of them on its own, which is what keeps a single coincidental match out of your report.

  5. 05 · REPORT

    It writes findings, and nothing else

    Masked samples, a confidence score, and the signals that produced it. The fidelity the scan actually achieved on each collection is recorded too, so "found nothing" and "could not look" are never confused with one another.

  6. AND THEN

    You revoke the key

    Nothing is left running. Lens is a batch job in a container rather than an agent living in your estate, so what you remove afterwards is a container.

How it decides

Three signals, and never one alone

A column called notes holding one stray number that happens to pass a checksum stays low confidence, while a column called aadhaar whose values also pass Verhoeff is near certain. Every finding carries the signals that produced it, so a reviewer audits the reasoning rather than trusting a score.

SIGNAL 01

The field is named

Checked against a hand-editable dictionary of aliases, so a column called uid_no is a candidate before a single value is read. Your own naming conventions go into the same dictionary, which is a YAML file rather than a code change.

SIGNAL 02

The values have a shape

A pattern applied to normalised values across the whole sample and scored as a hit rate, not as a single match. Twelve of the nineteen identifier types carry a pattern; the other seven are caught by name and context instead.

SIGNAL 03

The numbers check out

Aadhaar carries a Verhoeff check digit and GSTIN a mod-36 one, and where a published checksum exists Lens runs it. Never on its own: roughly one in ten random numbers of the right length passes, which is a false positive waiting to be printed.

  • Identifiers hiding inside prose

    A phone number in a remarks note is reported as a span inside the value with an occurrence count, not just as a whole-value match. Serialised JSON inside a text column is descended too, within bounds, so an identifier buried in a blob is still found.

  • Business data told apart from personal data

    A company PAN is not personal data under the Act, and the fourth character of a PAN says whether the holder is an individual. Lens reads that character, so a column full of vendor registrations does not land in your report as a pile of data principals.

What it can read

Adding a store is a file, not a release

Every source is a connector manifest in YAML rather than code, so a new member of a family is one file with no Python and no interface change. 32 sources and 4 findings feeds ship today. Hover or tap a mark to see what Lens reads there and what it has been tested against.

32 sources/4 findings feeds/read-only/findings out, never records

Run against a live instance Ships, tested against an emulator so far Findings feed, where results go
  • Managed databases need no separate connector

    RDS, Aurora, Cloud SQL and Azure Database are reached by the connector for the wire protocol they speak, and DocumentDB and Cosmos DB on the Mongo API by the MongoDB one. There is no connector per cloud product. What they do need is TLS configured against the provider CA bundle, which is documented per provider.

  • Documents, and photographs of documents

    PDF, Word, Excel, CSV and HTML are extracted by a bounded subsystem, and scanned images such as a photographed Aadhaar card or a signed form are read by optical character recognition. Extraction always records a reason when it declines to parse something, whether oversized, encrypted, malformed or a zip bomb, so a skipped file never looks like a clean read.

What comes out

A findings document, or a RoPA you still own

FORMAT 01

The findings document

Every detected field with its type, confidence, signals and a masked sample, plus the fidelity the scan achieved on each collection. Written to disk with zero outbound traffic, so you hand it over or upload it yourself. The same document can be posted to a control plane you name instead, if you would rather that was automatic.

FORMAT 02

Draft records of processing

The same scan, shaped as draft records of processing for the Sammati consent manager: one draft per source, detected categories arriving with their evidence and confidence, and the system identity filled into Where Stored. Purpose, legal basis, retention, recipients and notices stay blank, because those are the fiduciary's judgments and not a scanner's guesses. An existing findings file can be reshaped into this without rescanning anything.

What it does not do yet

Stated plainly, so nobody demos something that does not exist

A buyer who finds an unstated gap stops believing the stated claims too. These are the ones you would otherwise discover in week three of an evaluation.

No accuracy figure, anywhere

There is no hard-coded number in the product and there is none on this page. The evaluation harness, run against synthetic labelled fixtures, is the only source of a figure, and a number measured on somebody else's data shapes would mislead you about your own. We run the harness for your shapes during an assessment and show you the output rather than a claim.

A bare name in a sentence needs the optional model

Labelled and cued cases are caught deterministically, so a date of birth written as "DOB: [date_of_birth]" and a condition written as "diagnosed with" are both found. An unlisted name sitting loose in prose needs the optional model ensemble, whose weights are vendored into the image and loaded from local files only. A model-derived finding carries its own signal, so a reviewer always knows which claim came from where.

Some sources have only met an emulator

SQL Server, DuckDB, DynamoDB, OpenSearch, Snowflake, Redshift, Amazon S3 and Azure Blob Storage ship today, but so far each has been run against an emulator or a local server rather than a live account, and the source map marks them so. The same is true of the Splunk and Microsoft Sentinel feeds. Elasticsearch is reached through the OpenSearch connector without having been tested against it, and Cassandra is a manifest nobody has written yet.

The licence is backed by contract first

Lens makes no outbound calls by design, so expiry and sweep counting are checked offline against the clock of the host it runs on. The signature itself cannot be forged and the limits inside a licence cannot be altered, and a host owner who rolls the clock back and restores a snapshot can still reset the budget. We say so rather than describing a compiled image as secure.

How it is bought

Once, or continuously

A DPDP gap assessment

One engagement, a bounded number of sweeps, and a findings document you keep. This is the right shape when the question is what you actually hold, which is the question almost everybody starts with.

Continuous monitoring

Scans through a paid period, so a new column of Aadhaar numbers in a new service shows up in weeks rather than at the next audit. Findings can feed the records of processing in the Sammati consent manager directly.

There is no published price and no self-service tier, for the same reason the vault has none: the number depends on how many systems come into scope, and every engagement so far has needed a conversation before a figure would have meant anything.

Questions we get asked

Before you ask your security team

Does Lens send our data anywhere?

Lens makes no outbound calls to any hosted API, ours included. Detection is deterministic: regular expressions, hand-maintained dictionaries and published checksums. The optional model runs from weights vendored into the image and loaded from local files only, never fetched at runtime, so a perimeter with zero egress still works. The only thing that leaves is the findings document, and you are the one who moves it.

Do we have to tell it where the personal data is?

Discovery is the job of the tool rather than yours. You give Lens a connection and name a database, and it enumerates and samples everything in that scope by itself, including fields nobody documented and serialised JSON hiding inside text columns. Pointing a scanner at the places you already suspect is how an inventory ends up confirming what somebody assumed.

What accuracy should we expect from Arbor Lens?

No accuracy figure is published for Arbor Lens, deliberately. A number measured on synthetic fixtures or on another customer's data shapes would mislead you about yours, and it would then outlive the scan it came from. The evaluation harness reports precision and recall per identifier, and we run it against your shapes during an assessment, so the number you get describes your data.

What credentials does Lens need, and does it write anything?

Read-only credentials are enough, and temporary ones are ideal. Lens issues no writes, and when a scan completes it reminds you that the credentials can be revoked. A credential can also stay out of the configuration file entirely, by pointing at an environment variable or a mounted secret instead, and those references are resolved only for a config file and never for a connection somebody types into the console.

Can Lens scan files rather than databases?

Files are supported, on the host and in cloud storage. Arbor Lens reads a local directory tree, Amazon S3 and S3-compatible stores, Azure Blob Storage, Google Cloud Storage, SharePoint, OneDrive and Google Drive, plus mail attachments in Exchange Online and Gmail. PDF, Word, Excel, CSV, HTML and scanned images are extracted and scanned, with optical character recognition for photographs of documents. Amazon S3 and Azure Blob Storage have so far been tested against emulators rather than live accounts.

How does Lens relate to the Sammati consent manager?

Lens answers what you hold, and the Sammati consent manager answers whether you may hold it. The two are sold and deployed separately and neither requires the other. The one place they meet is the findings export, which can arrive as draft records of processing in the consent manager with every judgment field left for you to fill.

Is Lens available as a hosted service?

Arbor Lens is not offered as a hosted service, and that is the design rather than a roadmap item. A discovery tool reaching into your databases from outside your perimeter is a worse risk than the problem it solves, because it becomes a second copy of your data in a risk register you do not control. Lens runs where your data already is, which means we cannot reach it, cannot be compelled to produce it, and cannot lose it on your behalf.

How much does Arbor Lens cost?

Arbor Lens is quoted per engagement. The figure depends on how many systems come into scope rather than on how many records they hold, and a one-time gap assessment is priced differently from a continuous monitoring subscription. There is no published price and no self-service tier.

Last reviewed . General information, not legal advice. Lens is a separate product from the Sammati consent manager and from Arbor Vault, and all of them are listed here.

Where to start

Bring us the one database nobody wants to own.

An assessment runs inside your perimeter, against your own systems, with credentials you revoke afterwards. It ends with a findings document you keep, whether or not you buy anything else.