The DataSecOps Audit: Twelve Checks With Thresholds

The DataSecOps Audit: Twelve Checks With Thresholds

An audit is only as good as its pass criteria, and most internal data-protection reviews fail before they start because “is our data secure?” has no measurable answer. This document is the fix: twelve concrete checks for a DataSecOps programme, each with a pass threshold or a measurement method, followed by the maturity scale that turns the results into a level and the 30/60/90 calendar for fixing what fails. It is written to be executed — print it, assign owners, schedule the ninety days. For readers new to the label: it is Satori Cyber’s August 2021 coinage for an agile, security-embedded operating style around changing data and users — a vendor’s term, not a ratified standard, but the checks below grade the practice, not the vocabulary. Run them as written, keep the evidence each one produces, and the next external audit becomes an exercise in filing rather than firefighting.

Scoring: the DataSecOps maturity scale and its gates

The Ops Atlas maturity ladder for DataSecOps counts five stages — starting at ad hoc, through classified and enforced, then evidenced, ending at adaptive — and two gates carry most of the meaning. Crossing into the enforced level requires protection on more than half of tier-1 tables. The evidenced level demands better than 90 percent coverage plus zero console-only policies — everything versioned, nothing hand-configured. The final stage, adaptive, has a behavioural test rather than a numeric one: a newly registered sensitive table gains protection before anyone is granted access to it, every time, without anyone chasing it. Grade honestly against the gates; a programme that reports level three while console policies still exist is filing fiction.

The twelve checks

  1. A classification standard exists as a written document. Pass: the document names what counts as IDENTIFIER and SENSITIVE, who confirms scanner output, and how disputes resolve. Fail: classification lives in one engineer’s head.
  2. Discovery has run within the last quarter. Pass: a sweep of the estate with dated output listing every table holding sensitive fields, including copies and snapshots. Stale discovery is how ghost databases survive.
  3. The top-50 sensitive tables are fully classified. Pass: 100 percent of the list tagged, human-confirmed. This is the 30-day gate in the remediation calendar below.
  4. Tier-1 policy coverage above 50 percent — then above 90. Pass at the enforced gate: more than half of tier-1 tables carry query-time protection. Pass at the evidenced gate: better than 90 percent. In a worked example estate, protected columns stood at 1,085 of 1,240 — an 87.5 percent ratio that clears the second gate.
  5. Zero console-only policies. Pass: every masking and row rule traceable to a reviewed commit. Any rule configured by hand fails the check, because hand rules cannot be restored, diffed or evidenced.
  6. Multifactor authentication on every human principal. Pass: 100 percent. The 2024 tenant-breach campaign taught the price of skipping this one.
  7. Plain-text read share measured and declining. Measure: of all reads against sensitive columns, the fraction returning unprotected values. The example estate: 17.1 percent. Pass: the number is computed at all and its trend line points down.
  8. Over-exposed principals counted and shrinking. Measure: principals holding access beyond their role’s need. The example: 212 of 940 principals, 22.6 percent. Pass: measured quarterly, with a named owner for the reduction plan.
  9. Dormant grants surfaced every cycle. Measure: grants unused for 90 days. The example estate: 310 out of 1,900 grants — 16.3 percent. Pass: the list is produced and revocations actually follow. A dormant-grant list nobody acts on is shelfware.
  10. Access requests carry purpose and expiry. Pass: a request specifies dataset, purpose and duration; the owner approves; the grant expires automatically. Standard requests same-day; exceptions require a filed ticket, documented risk acceptance and a mandatory end date.
  11. The quarterly review revokes. Pass: each review produces at least one revocation, or a documented reason none was warranted. Reviews that never revoke are the ritual the audit exists to catch; the give-away is a review record whose revocation column has stayed empty for years while staff turned over twice.
  12. Non-production runs masked. Pass: lower environments rebuilt with masked data rather than cloned production. The developer sandbox holding raw production rows is the classic finding everyone claims to have fixed.

The 30/60/90 remediation calendar

Failing checks need sequencing, not panic. Day 30: classification standard written, discovery sweep executed, top-50 tables classified by hand, MFA rolled out to every human account. Day 60: tag-based policies live on tier-1 tables, the access-request workflow operating with expiries, a drift check running in the pipeline to catch untagged copies. Day 90: the first quarterly review completed with at least one revocation, non-production rebuilt on masked data, and the first scorecard — the twelve checks above, scored — circulated to stakeholders. Three months from zero to evidence-grade is the realistic tempo for a focused team. Programmes that shuffle the order — buying tooling on day one, writing the standard on day sixty — arrive at day ninety with software installed and nothing defensible to show for it, because the evidence chain depends on each phase feeding the next.

Evidence discipline: how to record a DataSecOps pass

A check without artefacts is an opinion. The evidence set per check: the dated document (checks 1), the sweep output (2), a tag-coverage export (3, 4), the repository listing showing zero console rules (5), the IdP MFA report (6), and for the measured checks (7-9), the query or script that computed the number plus its output, run quarterly. Two properties make evidence audit-grade: reproducibility — anyone can rerun the computation — and provenance — the artefact names when it was produced and by what. Scorecards assembled from screenshots fail both.

  • Score against gates, not vibes: >50 percent coverage is a level; 90 percent plus zero console rules is the next one.
  • Four numbers to trend quarterly: policy coverage, plain-text read share, over-exposed principals, dormant grants.
  • Reviews must revoke; a review with zero revocations needs a written justification.
  • Three months, three phases: classify, enforce, evidence — in that order, never shuffled.

Teams that want the surrounding pipeline discipline will find the wider field mapped in our guide to common DataOps challenges in analytics projects; the structural contrast with delivery automation is drawn in DataOps vs DevOps: differences and principles; and the automation practices that keep the drift check honest are covered in pipeline automation practices every engineer should know. The maturity ladder and its thresholds are documented in full at the Ops Atlas DataSecOps reference, which this checklist mirrors in operational form.