SCA, SAST and DAST Best Practices for 2026: Count Fixes, Not Findings
Equifax had a scanner. Your team probably has three. A triage command, a gating table and a 30-day rollout that turn SCA, SAST and DAST alerts into merged fixes.
Key takeaways
- Scanners pay off only when findings turn into fixes.
- Rank dependency findings by CISA KEV, EPSS and reachability instead of CVSS, block pull requests only on new findings that have a fix, and run authenticated, spec-driven DAST in staging on every release.
- Measure time to remediate and fix rate, not finding counts.
On this page
Your scanners are not the problem. Equifax had one. Most teams fail on routing: findings ranked by the wrong number, gates that block on old debt, and alerts that land on nobody. Rank by exploitation evidence, block only on new findings with a fix, give every finding one owner, and count fixes instead of findings.
Verified against
Against GAO-18-559, CISA advisory AA23-158A, CISA's XZ Utils and Log4j guidance, the CISA KEV JSON feed (catalog 2026.11.13, 1,781 entries), the FIRST EPSS API (scores dated 2026-11-15) and the FTC's 2018 Uber settlement. We ran the triage command below on November 16, 2026.
Equifax had a scanner. Nobody owned what it missed
According to the GAO's 2018 report, US-CERT published the Apache Struts remote-code-execution flaw in March 2017, and two days later, on March 10, attackers were already probing Equifax for it. The patch notice went to an out-of-date recipient list, so the dispute-portal owners never saw it. A network scan a week later missed the vulnerable server. Attackers came back on May 13, stayed about 76 days, and took data on at least 145.5 million people.
No tool was missing. The inventory was stale and the alert reached nobody.
Triage in one command: KEV and EPSS before CVSS
CVSS says how bad a flaw could be. CISA's KEV catalog says attackers are exploiting it now. FIRST's EPSS estimates the probability of exploitation in the next 30 days, rescored daily. Both are free. Paste in the CVE IDs from any scanner's output:
curl -s -o kev.json https://www.cisa.gov/sites/default/files/feeds/known_exploited_vulnerabilities.json
CVES="CVE-2021-44228,CVE-2023-34362,CVE-2017-5638,CVE-2024-3094"
curl -s "https://api.first.org/data/v1/epss?cve=$CVES" | jq -r --slurpfile kev kev.json \
'.data[] | .cve as $c | [$c, (.epss|tonumber*100|floor|tostring+"%"),
(if ($kev[0].vulnerabilities | any(.cveID==$c)) then "KEV" else "-" end)] | @tsv'
CVE-2024-3094 90% -
CVE-2023-34362 99% KEV
CVE-2021-44228 99% KEV
CVE-2017-5638 99% KEV
The xz-utils backdoor scores 90% on EPSS but is not in KEV: EPSS predicts, KEV confirms, so check both. Then add the two facts only you know: is the vulnerable code reachable, and is the service exposed to the internet? Our starting policy:
| KEV-listed | Reachable | Internet-facing | Action |
|---|---|---|---|
| Yes | Yes | Yes | Incident. Fix or mitigate within 24 hours |
| Yes | Yes | No | Fix within 7 days |
| No, EPSS high | Yes | Any | Fix within the sprint |
| No | No | Any | Backlog, batched into routine updates |
SCA: automate the boring fixes, distrust new packages
Decide the merge policy before turning on the update bot. Patch and minor bumps that pass tests merge automatically. Majors get a named reviewer within a week. Brand-new releases wait out a cooldown of a few days, so a hijacked release gets reported and pulled before it reaches you.
Keep an SBOM for every build. Log4Shell (CVE-2021-44228, December 2021) was an inventory test: CISA told organizations to enumerate every asset, including in-house software. With stored SBOMs, "where do we run log4j-core 2.14?" is a query. Run license policy off the same SBOM: allowlist passes, denylist blocks, unknown goes to review.
Defend against packages that are malicious on purpose. CVE matching cannot flag them. In event-stream (2018), a volunteer maintainer took over and added flatmap-stream, whose encrypted payload targeted cryptocurrency wallet software. xz-utils (CVE-2024-3094, March 2024) shipped a backdoor inside its official 5.6.0 and 5.6.1 releases. AI assistants add a third route: a USENIX Security 2025 study found commercial models invent package names at least 5.2% of the time and open-source models 21.7%, and anyone can register those names. Check every new dependency for existence, age and known-malicious status before merge, as our CI gate for AI-hallucinated packages does.
SAST: scan the diff, tune per rule, baseline the past
Report only what the pull request changed. A full-repo report buries two new findings under four hundred old ones, and people learn to scroll past the check. Snapshot existing findings on day one, gate only on new ones, and burn the baseline down separately with a target date. A gate that is red on day one gets disabled on day two.
Tune per rule, not per tool. Track each rule's false-positive rate. A rule that is wrong more often than right gets rewritten to recognize your sanitizers, or switched off. Every suppression carries a reason and an expiry date.
Catch secrets before the push. In the 2016 Uber breach, the FTC found that an engineer had posted an access key on a code-sharing site. Intruders used it to download files holding more than 25 million names and email addresses and 600,000 driver's license numbers. Scan in pre-commit and at push. If a secret lands anyway, rotate it first; rewriting git history does not un-leak it.
Treat AI-generated code as untrusted input. Agent-authored pull requests get the same gate with no exemption. Weight rules toward the classics: string-built SQL, disabled TLS verification, wildcard CORS. Our SAST vs DAST comparison covers what static analysis can't see.
DAST: authenticated, spec-driven, every release
Log in. An unauthenticated scan mostly tests the login page. Script authentication with one test account per role so the scanner can test whether user A can read user B's objects.
Scan the API from its OpenAPI spec. A crawler never finds endpoints no page links to. MOVEit Transfer (CVE-2023-34362) was SQL injection in an internet-facing web app, exploited by the CL0P ransomware gang from May 27, 2023. That bug class is what DAST exists for, though we won't claim any given scanner would have caught that flaw.
Run it on every release. A quarterly scan reports bugs that shipped up to three months earlier. Gate deploys on a time-boxed baseline scan and run the deep scan nightly.
Correlate. A DAST-confirmed injection plus the SAST finding in its handler is one ticket with proof of exploitability, and it goes to the top of the queue.
Where each one runs and who owns the fix
| Stage | What runs | Blocks on | Fix owner |
|---|---|---|---|
| IDE | SAST, secrets, new-dependency check | Nothing, warns | Author |
| Pre-commit and push | Secrets detection | Any verified secret | Author |
| Pull request | Diff SAST, SCA, license | New high or critical with a fix | PR author |
| CI build | Full SCA, SBOM, malicious-package check | KEV-listed or known-malicious | Owning team |
| Staging | Authenticated DAST, API scan | Confirmed high or critical | Owning team |
| Production | SBOM matched against new CVEs and KEV | Never blocks, pages per the triage table | Owning team |
AppSec owns rules and triage quality; product teams own fixes. A finding mapped to no team never gets fixed.
A 30-day rollout that survives contact with engineers
- Week 1: map every repository and service to an owning team. Turn on secrets blocking at push.
- Week 2: snapshot the SAST and SCA baselines. Switch pull-request checks to new findings only, in warn mode.
- Week 3: run the KEV and EPSS triage on the backlog and fix every reachable KEV item. Turn on blocking for new, fixable findings.
- Week 4: add authenticated DAST against staging on each release, then report the first MTTR and fix-rate numbers.
Seven rules worth printing
- Never break the build on a finding with no available fix.
- Block on new findings, never on the baseline.
- A suppression without a reason and an expiry date is a bug.
- If a rule is wrong more often than right, turn it off today.
- Rotate a leaked secret first, clean history second.
- No exemptions for AI-authored pull requests.
- One finding, one ticket, one owner, however many scanners reported it.
Metrics that matter
- MTTR for criticals: median days from detection to fix deployed, split KEV from non-KEV.
- Fix rate: findings closed by a code or version change, divided by findings opened. Suppressions do not count.
- False-positive rate per rule: this number drives tuning.
- Auto-remediation share: fixes merged from automated pull requests, divided by all fixes.
Stop reporting "findings found" to leadership. It rewards the noisiest scanner.
Questions people ask
Should SCA, SAST or DAST findings break the build?
Only new findings that have a fix and clear an exploitation or severity bar: new high or critical SAST findings on changed lines, KEV-listed or known-malicious dependencies, and confirmed high-severity DAST results in staging. Everything else warns and becomes a ticket with an owning team.
Is CVSS enough to prioritize vulnerabilities?
No. CVSS rates potential severity, not likelihood. KEV confirms active exploitation, EPSS predicts it over 30 days, and reachability tells you whether it applies to your code. A reachable KEV finding outranks an unreachable CVSS 10.
The call we'd make
Detection is a solved problem. Routing is not. Equifax's notice went to a stale mailing list, and the same thing happens today when findings land in a dashboard nobody owns. Fix ownership and gating first, and tune the scanners second. One question for you: which gate do your engineers bypass most often, and what would it take for them to stop?
Sources
Cite this article
Kiril Urbonas (2026). SCA, SAST and DAST Best Practices for 2026: Count Fixes, Not Findings. DevOpsNess. https://www.devopsness.com/blog/sca-sast-dast-best-practices-2026
[SCA, SAST and DAST Best Practices for 2026: Count Fixes, Not Findings](https://www.devopsness.com/blog/sca-sast-dast-best-practices-2026) (Kiril Urbonas, DevOpsNess, 2026-11-16)
More from DevOps
Explore more articles in this category
PHP2Shell (CVE-2026-109618, CVE-2026-109619): Patch Before the Bots Do
Two unauthenticated RCEs in PHP core, leaked a day before the patch, with exploit attempts already logged. Here is how to tell if you are exposed and what to do tonight.
Docker Security Best Practices: Images, Runtime, and Supply Chain
Harden container images and runtime. Image scanning, minimal base, and supply chain security.
How to Undo the Last Git Commit (Every Case)
You committed too early, wrote a bad message, or included the wrong file. Here's the decision tree for undoing a Git commit safely, by what you actually want to keep.
You might have missed
Evergreen posts worth revisiting.