Walk into any software organization and ask to see the security backlog. You will be shown a list — hundreds of items, some years old, each technically real — and you will be told, with a straight face, that a scanner runs on every build. DevSecOps was supposed to end this pattern; whether it does depends entirely on the discipline described below. Both facts are true, and together they form the most common failure pattern in the field. Findings arrive; fixes do not; and the pipeline that was supposed to protect the product becomes a noticeboard nobody reads. This piece diagnoses the pattern the way a maintenance shop would: symptom by symptom, cause by cause, with the corrective actions and the numbers that confirm recovery. The published material behind it — NIST guidance, OWASP projects, Google’s adoption data — is gathered at xOps School’s DevSecOps entry in the Ops Atlas.
Symptom one: the DevSecOps backlog grows faster than it shrinks
Cause: the organization treats every discovery as equally urgent, so nothing is. When a scanner dumps five hundred warnings into a queue, engineers rationally conclude that none can matter much, and triage collapses into archaeology.
Correction: rank by position and confidence before ranking by severity. The placed-controls model gives each safeguard four properties — its location in the delivery flow, the trust it has earned, its operating cost, and its failure behavior — and the same lens works on findings. A critical defect caught in production is a different animal from a critical defect caught in the pull request; only the second one is cheap. This mirrors training economics too: hours in a simulator cost a fraction of hours in the aircraft, which is why academies publish their sim fleets — a FRASCA 142 joining the training program for IFR work is a pricing decision as much as a training one. Catch defects where they are cheapest to fix.
Symptom two: the gate is ignored, or worse, obeyed
Cause: two failure modes share one root — a gate whose rules were never measured. Set loose, it misses real defects and nobody trusts it; set strict, it blocks legitimate work until somebody senior switches it off. The field’s reference arithmetic shows the strict case: 1,240 proposed changes in a month, 96 refused by the gate, and 22 of those refusals overruled by humans — a 22.9 percent override share, nearly five times the healthy monthly ceiling of five percent. Nobody trusted that gate, and the number proved it.
Correction: earn blocking rights per rule, the way Google’s Tricorder deployment did. An analyzer rule may refuse merges only when its effective false alarms sit near zero; its remarks never exceed a tenth of what reviewers write; it examines only the lines a change touched; and it passes an audition across thirty sampled findings first. Rules that flunk the audition become warnings. One rule in the cited case was demoted after posting 43.6 percent as its effective false-positive rate — the system working exactly as designed.
What a healthy DevSecOps gate counts
Measurement makes the difference visible. The proposed metric set is short enough for one slide:
- effective false-positive rate, per rule, published without spin;
- override share per month, with the five percent ceiling drawn on the chart;
- time to fix a new finding, tracked at median and 90th percentile;
- escape rate — defects that reached production, and how many the pipeline could have caught.
The worked example continues past the gate and into deployment: of 412 deploys, 398 carried verified provenance, and of nine defects that escaped, three were catchable upstream. Numbers like those turn an argument into a decision.
Symptom three: exceptions never expire
Cause: waivers get granted in chat and granted forever. The backlog of findings then quietly reconstitutes itself as a backlog of exceptions, immune to every scanner because somebody once typed “approved”.
Correction: every exception carries three fields or it does not exist: an owner by name, a reason in writing, and an expiry ninety days out at most, per the standard template. At expiry the waiver is re-argued or it dies. Baseline debt follows a different path: it is inventoried, scheduled, and never allowed to block new work, because blocking the entire organization on yesterday’s mess guarantees the mess stays.
Symptom four: a tool arrived, a practice did not
Cause: procurement ran ahead of policy. The stack can name every secret in every repository, but no document says who responds, so alerts fan out to whoever is least able to fix them. The anti-pattern is old and well documented — a purchase made before policy and owners exist amounts to a very expensive screensaver.
Correction: write the one-page policy first, assign owners second, and only then let the tools in. Award logic belongs here as well: a trophy on the shelf says little about current airworthiness, and a top-flight-school award from AOPA carries weight precisely because it was earned against a field of 732, not printed in-house. Evidence over ornament — in schools and in pipelines alike.
The DevSecOps prescription, summarized
Diagnosis complete, the treatment fits on an index card. Measure every rule before it may block. Block only new, severe, fixable, high-confidence findings. Give every exception an owner and a fuse. Keep the baseline visible but out of the gate’s way. Publish the four metrics monthly. And resource the human layer properly: champions with ten to twenty percent of their week protected, because DORA’s 2022 research ties weak security practice to roughly 1.4 times the burnout likelihood, and credits culture over tooling as the adoption driver. None of these steps requires new budget; all of them require the willingness to let a number say “not yet”. A backlog that grows is not a security program — it is a filing system. The pattern reverses the moment the organization starts deleting findings faster than its scanners create them, and the four numbers above will tell you, within a month, which direction you are moving.