Glossary. Sensitive data discovery.
What is sensitive data discovery?
Discovery finds where sensitive data sits and records what the scan couldn’t read. Without that second half, a “nothing found” can’t be trusted.
By Dipal Panchal, co-founder and CEO · Updated November 8, 2026
Where the data sits, and what wasn’t read
Sensitive data discovery is the process of scanning databases, files and cloud services to find where personal, financial, health or credential data sits, down to the table, column or file. Unlike data classification, which sets the level each finding takes and the handling that follows, discovery answers where the data is and records what the scan couldn’t read. A finding count means little until the report says what was searched.
Blind spots start at the listing step
A discovery scan connects to each source with read access and first lists what is there. That means schemas, tables and columns in a database; buckets, folders and files in cloud storage; sites and libraries in Microsoft 365. Anything the scan account can’t list is missing from that list, and no later step can report it.
For each item it can list, the scan reads a sample of values and tests them against patterns, such as the shape of a card number. A pattern also matches look-alikes. Where the identifier carries a check digit, a match shouldn’t count as a finding until the check passes; where there is no check to run, a person confirms it first. Each finding is saved with its location and scan date, and each listed item the scan couldn’t read is saved with the reason.
The Luhn check, digit by digit
Payment card numbers end in a Luhn check digit. Here it is on a published test card and on a 16-digit shipment reference, computed with a short Python 3 script on November 7, 2026.
| Step | Test card 4111 1111 1111 1111 | Shipment reference 7700 1826 3341 0592 |
|---|---|---|
| 1. From the right, double every second digit; subtract 9 from any result over 9 | 8 1 2 1 2 1 2 1 2 1 2 1 2 1 2 1 | 5 7 0 0 2 8 4 6 6 3 8 1 0 5 9 2 |
| 2. Add the 16 digits | 30 | 66 |
| 3. Divide by 10 | Remainder 0: passes | Remainder 6: fails |
| 4. What it means | Card-shaped and checksum-valid; context decides the rest | Not a valid card number; the scan drops it |
Six results from four markets
A generic scan, not IRIS output, that runs discovery and classification in one pass: Location is the discovery answer and Sensitivity the classification one.
| Source | Location | Detected type | Why | Sensitivity | Decision |
|---|---|---|---|---|---|
| HR database, UK | hr.employee.ni_number | UK National Insurance number | NINO-format values whose prefixes pass HMRC’s allocation rules, in a column named ni_number | Restricted | Kept; HR confirms the column |
| HR database, UK | hr.employee_seed.ni_number | None (NINO candidate) | QQ 12 34 56 C: HMRC never uses Q as a prefix letter, so the rows are seeded placeholders | None | Rejected: test placeholders, not identifiers |
| Cloud storage, US | exports/settlements_2026_10.csv, column card_no | Payment card number | 16-digit values that pass Luhn and match no published test number, outside the payments database | Restricted | Kept; the owner is asked why the export exists |
| Microsoft 365, UAE | HR site > Visa renewals > eid_list.xlsx | Emirates ID number (candidate) | Format only: 15-digit values starting 784, the UAE’s ISO country code, in a column headed EID; no check digit run | Restricted, proposed | Held as a candidate; counted once HR confirms the file |
| Customer database, India | crm.customer.ref_no | Aadhaar number | Every sampled 12-digit value passes Verhoeff, in a column whose name gives no hint | Restricted | Kept; the CRM owner confirms the column |
| Cloud storage, US | contracts/scans/2024/, a folder of PDFs | Not read | Scanned images with no text layer, and no text recognition was run | Unknown | Excluded and listed as not read, never as clean |
Rows 2, 5 and 6: a bare pattern, a name search and an unrecorded gap
Row 2 fails a published rule: HMRC’s National Insurance manual says D, F, I, Q, U and V are never used as either prefix letter. A pattern-only scan would report a column of placeholders as identifiers.
Row 5 is the reverse: a column called ref_no would slip past a name-based search, and the check digit found it. Random 12-digit numbers pass Verhoeff about one time in ten, and the evidence is the share of sampled values that pass, as with Luhn.
Row 6 is a folder the scan found and couldn’t read. Its true count is unknown, which isn’t the same as zero, and a scan that didn’t record unread items would have reported it as clean.
Where a finding stops
A passing check proves a value is well formed. It doesn’t prove the value belongs to a real person, and a test card number passes Luhn as readily as a live one.
Discovery also covers only what it read. A source nobody connected, a table the scan login can’t see, a file it couldn’t open and the rows outside a sample are outside the finding.
Count the tables the scan login can see
SQL Server and MySQL show a login only the tables it holds a permission on. Run this once as the scan login and once as an administrator; any difference belongs in the not-read count.
-- Read-only. Run in the same database, once as each account.
-- SQL Server counts every schema in the current database.
SELECT COUNT(*) AS tables_visible
FROM INFORMATION_SCHEMA.TABLES
WHERE TABLE_TYPE = 'BASE TABLE';
-- On MySQL, add before the semicolon: AND TABLE_SCHEMA = DATABASE()
-- Example output (synthetic):
-- administrator tables_visible = 214
-- scan login tables_visible = 187
-- The 27 tables the scan login can't see go in the not-read count.Report the denominator before the findings
Dipal Panchal, Sylox’s co-founder and CEO, puts the starting point bluntly. In October 2026 he said, “You can’t do anything else if you don’t know what sensitive data you have and where it is.”
Our rule: a discovery report opens with its denominator. That means the sources in scope, how many are connected, how many were scanned and what couldn’t be read, with the reason. In a September 2026 demo Panchal called a source that is connected but never scanned “a red flag” in its own right.
IRIS reads a sample of each table, and small tables in full, in the sources you connect, and dates every scan.
Locations down to the table, column or file: IRIS produces part of this, in the sources you connect.
Check digits on identifiers that carry one: IRIS produces part of this, for nine identifier types validated by checksum, such as cards (Luhn) and Aadhaar (Verhoeff).
What it couldn’t read, with the reason: IRIS produces part of this, for the tables and files the scan account can list.
Whether a copy should exist: Not produced by IRIS.
Explore discovery and classificationWrite down what you expect before you scan
Pick one source and note, before the scan, which sensitive types you expect it to hold and which tables or folders you’ll leave out. Afterwards, every expected type should show up as a finding or as an explained absence, and every area left out should be listed as not read.
The record that holds that denominator is scan coverage. If the source you pick is a copy of production, how to find production personal data in your test databases covers that case.
What a finding leads to
What is data classification?
The decision after a finding: which level it takes and the handling that follows.
What is shadow data?
What a finding outside the systems you track points to: exports, restores and extracts with no owner.
What is a sensitive data inventory?
The next step: the dated list of where sensitive data sits, built from discovery and kept by owners.
What teams ask before a discovery scan
Not usually. A sample supports a decision about the tables it tested, so ask any tool to state how it samples and which tables it read in full.
Yes, read access to wherever the data lives, production included. A current replica or snapshot can stand in if it holds every table the primary does and the report names it as the source. A logical replica holds only the tables it subscribes to.
After each new source is connected, after each fix and on a regular schedule. A finding describes the source on its scan date, and a dated re-scan after a fix is the record that the fix held.
Protect your brand. Start with your data.
Choose one database to start. We’ll show you what IRIS finds there and what it couldn’t read.