Cookie settings

Necessary, always on: they keep the site working and remember your cookie choice.

Skip to content
Skip to content

Guide. GDPR. Article 32.

GDPR Article 32: evidence for security of processing

Map each Article 32 measure to what an auditor tests, and check the stored form of your sensitive data against the systems that hold it.

Dipal Panchal

Co-founder and CEO · Updated November 8, 2026

On this page

Four kinds of evidence, one per measure

Article 32 of the General Data Protection Regulation (GDPR) asks controllers and processors for security appropriate to the risk. A policy that says data is protected is not evidence. Evidence is a dated record of how the data is stored, who can reach it and when someone last tested that.

The measures the article names each leave a different record, and the sections below take them in this order.

  • Article 32(1)(a): a check of the stored form of each sensitive column, and of the connection to it
  • Article 32(1)(b): who can reach each store, and when that access was last reviewed
  • Article 32(1)(c): the last restore test of the systems that hold personal data
  • Article 32(1)(d): dated test results with scope and method, re-run after each fix

Scope, roles and time needed

The SQL targets PostgreSQL 16, 17 and 18. Sylox ran them in its lab on versions 16, 17 and 18, on synthetic data. UK readers apply the same method under UK GDPR Article 32; the UK GDPR page covers the UK differences.

The database administrator (DBA) runs the queries and the data protection officer (DPO) or security lead reads the results against the risk assessment. One database takes about an hour. Collecting the same evidence from processors takes as long as they take to answer.

  • A list of the columns your classification marked as sensitive
  • A read-only role that can read those tables, and pg_read_all_stats for the connection query
  • Your latest risk assessment for the processing, since Article 32 is judged against it

What Article 32 requires

Article 32(1) asks for technical and organizational measures that give a level of security appropriate to the risk. The level takes account of the state of the art, the costs, the nature, scope, context and purposes of processing, and the risk. The list that follows is not exhaustive, and each measure applies as appropriate to the risk.

The four listed are (a) pseudonymization and encryption, which you check store by store, and (b) ongoing confidentiality, integrity, availability and resilience. The others are (c) restoring availability and access in a timely manner after an incident, and (d) a process for regularly testing, assessing and evaluating the measures.

Article 32(2) points the assessment at accidental or unlawful destruction, loss, alteration, unauthorized disclosure of and access to personal data. Recital 83 asks you to evaluate those risks and mitigate them. Check why a measure such as encryption fits the risk, because your risk assessment is part of the evidence.

Each measure, what the auditor tests, what you keep

The mapping is Sylox’s reading of the article. Your auditors judge whether a control is met.

Each measure, what the auditor tests, what you keep
Article 32 measureWhat the auditor testsEvidence to keepWhere in this guide
(1)(a) PseudonymizationWhether identifiers are separated from the additional information, and who holds itWhere the keys or lookup tables sit and who can reach themStep 2
(1)(a) Encryption checksWhether sensitive columns, copies and connections are encrypted, and who holds the keysQuery 1 and query 2 output, key custody recordSteps 1 and 2
(1)(b) ConfidentialityWho can reach each store, and the date of the last access reviewGrant export, owner decisions, review dateStep 3
(1)(b) Integrity and resilienceControls that detect unauthorized change, and capacity to recoverOutside this guide: use your backup and monitoring recordsNot covered
(1)(c) Restoring accessA restore test with a date, scope and resultThe restore test reportNot covered
(1)(d) Regular testingA repeatable test with dated results, re-run after fixesThe queries, their output and the re-runStep 4

Step 1: check the stored form of sensitive columns

Start from the data, not the policy. For each system that holds personal data, check whether the stored values are in plain, masked, hashed or encrypted form, and whether the copies of that data are too. A setting on the database tells you about the storage layer. A check of the values tells you whether the sensitive columns themselves are protected, which matters most for exports and extracts.

Query 1 samples three columns and reports the share of each form by pattern. A pattern query can’t tell strong encryption from random-looking encoding, and a format-preserving token reads as plain, so it narrows the review and doesn’t replace it. Check the glossary entry on data masking for how masking, tokenization, hashing and encryption differ.

Query 2 counts the client connections that use TLS and those that don’t. Article 32(1)(a) names encryption without saying where, so a column that is encrypted at rest and read over a plain connection leaves a gap an auditor will find.

PostgreSQL 16 to 18: stored form and TLS

Read-only. Replace the three columns with yours; the connection query needs pg_read_all_stats to see other users’ sessions.

SQLTested on PostgreSQL 16.15, 17.11 and 18.6, November 9, 2026, as the postgres superuser, against synthetic data.
-- Query 1: stored form of sampled values (patterns, not proof)
WITH sample AS (
  SELECT 'crm.customer.email' AS col, email::text AS v
    FROM crm.customer TABLESAMPLE BERNOULLI (5) WHERE email IS NOT NULL
  UNION ALL
  SELECT 'billing.payment_card.card_number', card_number::text
    FROM billing.payment_card TABLESAMPLE BERNOULLI (5) WHERE card_number IS NOT NULL
  UNION ALL
  SELECT 'hr.employee.national_id', national_id::text
    FROM hr.employee TABLESAMPLE BERNOULLI (5) WHERE national_id IS NOT NULL
)
SELECT col, count(*) AS sampled,
       count(*) FILTER (WHERE v ~ '[*xX#]{3,}')                              AS looks_masked,
       count(*) FILTER (WHERE v ~ '^[0-9a-f]{64}$' OR v ~ '^\$2[abxy]?\$')    AS looks_hashed,
       count(*) FILTER (WHERE v ~ '^[A-Za-z0-9+/]{32,}={0,2}$')              AS looks_encoded,
       count(*) FILTER (WHERE v !~ '[*xX#]{3,}' AND v !~ '^[0-9a-f]{64}$'
                          AND v !~ '^\$2[abxy]?\$' AND v !~ '^[A-Za-z0-9+/]{32,}={0,2}$') AS looks_plain
FROM sample GROUP BY col ORDER BY col;

-- Query 2: client connections with and without TLS
SELECT count(*) FILTER (WHERE s.ssl)     AS tls_connections,
       count(*) FILTER (WHERE NOT s.ssl) AS plain_connections
FROM pg_stat_ssl AS s
JOIN pg_stat_activity AS a ON a.pid = s.pid
WHERE a.backend_type = 'client backend';
SHOW ssl;

What query 1 finds, value by value

Five values from the sampled columns. The last is a rejected false positive: a pattern that looks hashed and isn’t.

Records
ValueDetected typeLabelReasonKept or rejected
jane.example@example.comEmail addressPlainMatches the address pattern; no masking or hashingKept
************1111Payment card numberFlagged as maskedAsterisks replace all but the last four digitsKept
a665a45920422f9d417e4867efdc4fb8a04a1f3fff1fa07e998e86f7f7a2e3a8Email addressFlagged as hashedSixty-four hex characters; unkeyed, so guessable for short inputsKept
4111 1111 1111 1111Payment card numberPlainPublished test number; spaces between groupsKept
3f2a9c1e5b7d40a8b6c4e2f1a0d9c8b7a6e5d4c3b2a1f0e9d8c7b6a5f4e3d2c1Hex string flagged as hashedNot personal dataSixty-four hex characters, but the column holds build references; confirmed with the ownerRejected

Step 2: check pseudonymization and key custody

Article 4(5) defines pseudonymization as processing personal data so it can no longer be attributed to a specific person without additional information. That information must be kept separately, under measures that stop the attribution. Recital 26 says pseudonymized data that can be attributed to a person by using additional information is still personal data.

So the evidence is about the separation. Check where the lookup tables or keys sit, who can reach them, and whether any copy of the data still carries the direct identifiers. In the records above, the unkeyed hash of an email address fails that test, because anyone can recompute it for a guessed address.

Key management belongs in the evidence: who holds the keys, where they are stored and who can use them. Article 34(3)(a) shows the stakes. If encryption made breached data unintelligible to anyone not authorized to access it, you may not have to tell affected people about the breach.

Step 3: show who can reach the data

Unauthorized disclosure of and access to personal data are among the risks Article 32(2) asks you to weigh. An auditor asks who can reach each store of personal data, through which accounts, roles and groups, and when that access was last reviewed. Service accounts and broad groups need the same scrutiny as named users.

The method is in How to run a database access review, and the output of its queries is the evidence for this step. Article 32(4) adds that people acting under your authority process personal data only on your instructions, so keep the instruction and the training record with the access review.

Processors carry the duty too. Under Article 28(3)(c), a processor’s contract must require it to take all the measures Article 32 requires. Ask each processor for the same three records you keep yourself: stored form, access and tests.

Step 4: keep dated tests and re-test after fixes

Article 32(1)(d) asks for a process for regularly testing, assessing and evaluating whether the measures work. The evidence is the dated result, not the plan to test. Keep each test with its scope, method, date, result and what it did not cover.

After a fix, re-run the same queries on the same scope, so the record shows the condition before and after. Article 32(3) adds that an approved code of conduct or certification can help demonstrate that you meet Article 32(1). It supports the evidence and leaves the measures in place.

What these queries miss

Each gap can make a column look better or worse than it is. Note the ones that apply when you file the results.

What these queries miss
GapWhat to do about it
Checking encryption at rest below the databasePostgreSQL’s docs place this at file-system or block level; ask the platform team for the volume or provider setting
Format-preserving tokensQuery 1 reads a token that keeps the real value’s shape as plain; check with the team that built the tokenization
Small samplesCheck the sample size: a sample can miss a plain column in a mostly masked table, so raise it for high-risk columns
Backups, exports and replicasThey hold their own copies; run the checks on each, or record why not
Checking application-level encryptionA ciphertext column looks encoded, and the query cannot tell strong from weak; review the design with the owner
Checking whether an encoding is encryptionBase64 is an encoding, so treat looks_encoded as a lead to verify and not a result

The evidence folder for Article 32

File these together and lock the folder once signed.

The evidence folder for Article 32
Evidence itemProduced by, when and whereCompleteness checkSupports
The risk assessment the measures answerDPO, before the queries runDated, and names the processing it coversGDPR Article 32(1)
Query 1 and 2 output, with the account that ran themDBA, on the review date, on each storeEach sensitive column in the classification appears onceGDPR Article 32(1)(a)
Key custody recordSecurity lead, on the review dateNames who holds each key and where it is storedGDPR Article 34(3)(a)
Access review output and decisionsDBA and data ownersA decision for every row; none by the granteeGDPR Article 32(2)
Processor evidenceDPO, from each processorOne response per contract under Article 28GDPR Article 28(3)(c)
The re-run after each fixDBA, after the changeThe same queries and scope as the first runGDPR Article 32(1)(d)
With IRIS

IRIS checks the form of sampled values in each sensitive column, once the check is activated on your install. Your team closes the gaps in the source system. Capabilities vary by edition.

A check of the form of sampled values: encrypted, masked, hashed or plain: IRIS produces this.

A check of the encryption-at-rest setting of the database itself: IRIS produces part of this, on 9 database engines, beside the form found in sampled values.

Who can reach sensitive data, as of the scan date: IRIS produces part of this, where the source exposes permission data.

Judging cipher strength or key custody: Not produced by IRIS.

Applying protection in the source system: Not produced by IRIS.

A dated re-scan that records the new state beside the first: IRIS produces this.

Explore protection checks

Run the queries again on the next review date

The check that closes this guide is the re-run in step 4. Check whether a column moved from plain to masked or encrypted after your change: the second output shows it, and the two files make the dated record. If it did not move, the change did not reach the data.

This guide does not cover backup and restore testing, data protection by design under Article 25 or breach notification under Articles 33 and 34. For UK GDPR work, IRIS maps its evidence to those controls through its unified compliance framework. A scored readiness view for that work isn’t in v1; ask us if you need one. This page is general information, not legal advice. The next guide in the sequence covers the records behind it: GDPR Article 30 records versus the real data estate.

Official sources

  1. GDPR Article 32, security of processing, EUR-Lex. Opened November 8, 2026.
  2. GDPR Article 4(5), definition of the term, EUR-Lex. Opened November 8, 2026.
  3. GDPR Article 28(3)(c) and Article 34(3)(a), EUR-Lex. Opened November 8, 2026.
  4. GDPR Recitals 26 and 83, EUR-Lex. Opened November 8, 2026.
  5. UK GDPR Article 32, legislation.gov.uk. Opened November 8, 2026.
  6. EDPB Guidelines 01/2025, consultation version, European Data Protection Board. Opened November 8, 2026.
  7. Accountability toolkit: records management and security, ICO (Information Commission). Opened November 8, 2026.
  8. PostgreSQL 18: pg_stat_ssl, PostgreSQL Global Development Group. Opened November 8, 2026.
Share this guide
FAQ

Article 32 questions

Protect your brand. Start with your data.

Bring one database. We’ll show you which of its sensitive columns are stored in plain form, read-only.