Kestrel Wings is a five-person type-rating school renting two classrooms above a hangar, and until this past September it managed its data the way most schools of that size do: a shared drive, a spreadsheet of who-could-open-what, and faith. DataSecOps — the discipline Ben Herzberg set down in 2021 while leading research at Satori Cyber, built to secure data that keeps changing in the hands of people who keep changing — arrived at Kestrel the way it arrives at most places: late, and as the result of a small, fixable accident. This is the story of that semester, told as it happened, because the sequence teaches more than any checklist would.
September: the email that should have stayed internal
The office manager, Farida, sent a re-validation schedule to the chief flying instructor and attached the wrong file — the full student register, with licence numbers and medical expiry dates, instead of the one-page roster. The recipient was fine; the attachment should not have existed as a single unguarded file at all. The school’s incident response, in full, was an apology and a promise. But Farida did something less common: she wrote down every place the register existed. Fourteen. The scheduling database, a desktop export, two shared drives, a lease partner’s portal copy, a trainer’s tablet, and a folder optimistically named OLD.
October: finding and naming, the DataSecOps way
The school’s part-time IT contractor, Mateo, took Farida’s fourteen-item list and turned it into October’s work: discovery and classification. A scanning tool proposed labels across the database, and — this is the part worth copying — the scanner’s output was treated as a draft. Classifiers, including the well-known ones from Google and Microsoft’s Presidio project, are never fully accurate, so a human confirmed each field marked identifier or sensitive before anything was enforced. Licence numbers, passport scans, medical dates, payment details: tagged. The Ald folder got its contents interviewed, most of it retired. On 4 August 2021 the practice had been given its name; in October it got its first local application, one spreadsheet at a time.
November: DataSecOps rules on the database, with eyes open
Kestrel’s register lived in PostgreSQL, which turned out to be convenient: row-level security has been part of the engine since version 9.5, released in January 2016. Mateo wrote policies so instructors saw only their own students’ rows, and the lease partner’s account saw columns rather than whole tables. Two lessons landed that month. First, the referential-integrity loophole: checks that keep foreign keys honest run past row policies, which practitioners call a covert channel — presence of a record can leak even when contents cannot. Second, pgAudit’s honest limit: it cannot reliably account for superuser actions, so Mateo paired it with a second control rather than pretending the log was complete. Both discoveries went into the school’s one-page standard as warnings, which is how a five-person operation remembers what a consultancy would put in an appendix. Rules were stored as scripts, because a rule that exists only as clicks in a console cannot survive a migration, and hand-copied tables drop their tags on the way across. The partner portal’s copy, meanwhile, was replaced by a filtered view, so the portal never again holds a private stash of the register — the data stays home, and the portal asks questions of it instead.
December: passwords, portals and the hashing lesson
The lease partner’s portal had been “protected” by hashing student email addresses, a practice Mateo had once defended. Research for this project changed his mind. A deterministic hash always produces identical output from identical input, which means a table of hashed addresses can be matched against a table of hashed guesses — making it pseudonymisation, not anonymity, at best. That same month, the extortion wave tracked as UNC5537 was still working through warehouse customers via credentials stolen from staff devices; 79.7% of the hit accounts had already crossed paths with an infostealer, multi-factor authentication was missing everywhere, and certain passwords had gone four years without a change. Kestrel’s response fit in an afternoon: multi-factor on every account, a schedule for rotation, and the old password-only logins retired. Nobody at a five-person school believes they resemble a global platform’s customers — the numbers say otherwise about human habits.
January: the first review that revoked
The semester’s quiet milestone arrived in January: the school’s first quarterly access review. Eleven grants existed on paper; four belonged to people or purposes that had ended — a trainer who left in the spring, the portal account for a course completed in July, Mateo’s temporary elevated role, and the OLD folder’s ghost. All four were revoked, in a meeting that took twenty minutes and produced, by Farida’s account, the quarter’s most satisfying paperwork. Published studies of larger operations put numbers on what such reviews sweep out: one worked example counted 310 grants, of 1,900 in all, unused past ninety days — a 16.3% dormancy rate. Kestrel’s ratios were worse than that, before January. The review also adopted the request pattern going forward: any new access names its dataset, what it is for, and how long it should last; the owner signs off; the grant carries an expiry from birth.
February: what changed, in the end
The measurable residue of the semester fits on an index card:
- One written classification standard, one page long, dated and owned.
- A register under row and column rules, stored as scripts, not clicks.
- Multi-factor authentication on every account, rotation on a schedule.
- Fourteen storage locations reduced to five managed ones.
- A quarterly review on the calendar, with revocations as its success metric.
- A request habit — dataset, purpose, duration — replacing hallway favours.
The maturity ladder practitioners use climbs from ad hoc past classified and enforced up to evidenced, and its top rung — adaptive — is the one worth explaining at length: a brand-new sensitive table is protected before anyone is granted to it, every single time, without a meeting. Kestrel reached evidenced this year; adaptive is a habit, not a project, and the school’s next goal is to hold it through a full course cycle without slipping. The full map of the practice behind the story — the loop of discovery, classification, policy, enforcement, egress, history and review — is laid out in the Ops Atlas DataSecOps entry at xOps School, which is the reference Mateo bookmarked for the February staff meeting, and the one Farida now quotes in induction briefings for new instructors.
One last scene, because it explains why any of this stuck. In March, a new course brought a fresh intake, and the register grew a new table. Farida tagged it and gated it the same afternoon, from the standard, without asking Mateo. That is the whole story in one sentence: the discipline stopped being a project and became how the school simply works — the way Kestrel’s instructors long ago stopped discussing checklists and started using them. DataSecOps did not arrive at Kestrel as software or as a memo. It arrived as one email mishap, one honest semester, and the modest, permanent relief of knowing exactly who can see what, and for how long.
