KEV sets the order. A CVE on CISA's Known Exploited Vulnerabilities list goes to the front of the queue, and exposure sets the clock. EPSS ranks everything that isn't on the list. CVSS ranks nothing. It describes how much damage a successful exploit does, and that is all it describes.

Federal policy now runs the same way. CISA's BOD 26-04, issued on 10 June 2026, replaced CVSS-driven patch deadlines with four yes-or-no questions: is the asset publicly exposed, is the CVE in KEV, can attackers automate exploitation, and does exploitation give total control.⁠1 EPSS isn't one of the four. CVSS isn't either.

You need this if your team writes patch SLAs, sorts a scanner export, or reads "Critical" as "patch first." Three things changed this year. NIST narrowed what it enriches in the NVD (15 April). CISA moved federal deadlines off CVSS (10 June). EPSS shipped a new model (15 June). The first two change what your data looks like. The third changes what your dashboards plot. Status as of 8 November 2026.

Three signals, three questions

SignalQuestion it answersMaintained byRefreshedBlind spot
CVSSHow bad is a successful exploit?FIRSTConstant unless rescoredAttacker interest
EPSSWill exploitation activity be observed in the next 30 days?FIRST SIG; Empirical Security builds the modelDailyPast exploitation, by design
KEVHas CISA confirmed exploitation?CISAMost weekdaysAnything CISA hasn't confirmed

FIRST's CVSS v4.0 guide says the Base score measures severity and shouldn't be used alone to assess risk.⁠2 It adds that NVD and product maintainers typically publish only the Base score, so anything past Base is the consumer's job.⁠2 A score that folds in threat intelligence and asset data gets closer to risk, FIRST says, if you automate the work.⁠2

What changed in 2026

On 15 April, NIST said it will enrich CVEs in KEV first, with a goal of one business day, then CVEs in software the federal government uses and in critical software under Executive Order 14028.⁠3 Everything else is labelled lowest priority and may wait. NIST also stopped routinely adding its own severity score when the CVE's assigner already supplied one, and moved the backlog published before 1 March 2026 into a Not Scheduled category.⁠3 Submissions rose 263% between 2020 and 2025, and NIST says that drove the change.⁠3 Opinion If your SLA reads a CVSS band from the NVD, a growing share of new CVEs will arrive with the assigner's score or none from NIST.

BOD 26-04 revoked BOD 19-02 and BOD 22-01.⁠1 Reported Its deadlines follow the four questions. Three or more yes answers mean three days. KEV alone means 14 days. Public exposure alone, or automatability alone, means 60. Total control alone, or nothing at all, means the next scheduled upgrade or rebuild. KEV plus total control, plus either exposure or automation, adds a forensic triage requirement.⁠4

CISA analyzed one unnamed agency and found about 1% of its vulnerabilities in the three-day tier and more than 60% deferrable.⁠5 The agency ties the change to AI-assisted discovery and exploitation shortening the time between patch and exploit.⁠5 Tod Beardsley, CISA's former KEV section chief, doubts that three days is achievable across more than a hundred agencies.⁠5 The directive binds federal civilian agencies only, and CISA urges everyone else to adopt similar practice.⁠5

Since the directive, CISA has added 131 entries to KEV through 16 October. 108 carry three-day deadlines, 23 carry 14 days, and none carry the old 21-day window.⁠6 87 of the 131 require forensic triage.⁠6 KEV entries now arrive with the clock attached.

EPSS moved to version 5 on 15 June. Reported Empirical Security, which builds the model, says v5 scored 0.633 against 0.514 for v4 on the same 4 May 2026 data, a 23% gain on its own metric.⁠7 That is the maker grading its own work, and I found no independent evaluation of v5. Opinion Any EPSS trend line that crosses 15 June mixes attacker behavior with a model change.

Where each signal fails

CVSS

One bug can carry several Base scores. For CVE-2026-111497, the Brotli decoder flaw, Google scored the library 9.2 under CVSS v4.0, Node.js rated its own build High, and the .NET build is listed at 8.1.⁠8 FIRST's guide allows separate Base scores for different product versions and platforms,⁠2 so none of the three is wrong. None of them sets a queue order either.

The reverse case turns up too. On 3 November, Fluid Attacks' database showed CVE-2026-47208, a vm2 sandbox breakout, at CVSS v3.1 10.0, EPSS 1.04% (63.79th percentile), no KEV entry and no public exploit on record.⁠9 That is one vendor database on one day. It shows two scores disagreeing, which is what two different questions produce.

Measured In the EPSS team's own December 2022 test, patching everything at CVSS 7 or above meant acting on 58.1% of published CVEs to cover 82.1% of those exploited in the following 30 days, at 3.9% efficiency. An EPSS v3 cutoff of 0.088 covered 82.0% while acting on 7.3% of CVEs.⁠10 The authors build EPSS, the data is one snapshot, and the model tested was v3.

EPSS

EPSS was built blind to past exploitation. NIST's CSWP 41 explains the choice: adding past exploitation as an input made the model reactive, so the maintainers kept it out, and a CVE exploited this month gets no bump next month.⁠11 The EPSS FAQ, which NIST reproduces, says evidence of exploitation supersedes the score, and describes EPSS as working best "just below the obvious."⁠11 NIST's conclusion is that EPSS shouldn't be used alone.⁠11 That paper predates v5. Empirical's v5 notes list changes to optimization, calibration, exploit-code detection and data feeds, and don't mention past exploitation as an input.⁠7

The labels behind the model count any recorded exploitation attempt, whether or not it succeeded or the target was vulnerable. They come from signature-based sensors, which lean toward network attacks on perimeter devices.⁠10

FIRST treats the cutoff as a local decision, as Fluid Attacks reads its guidance. A 50% cutoff sounds balanced and deprioritizes 98.8% of published vulnerabilities.⁠9

KEV

CISA lists a CVE when it has an ID, reliable evidence of exploitation in the wild and a clear fix. Attempted exploitation counts, including attempts against honeypots. Scanning, research and public proof-of-concept code don't.⁠12 The list is small. As of catalog version 2026.10.16 it held 1,748 CVEs, and its 189 entries from 2025 compare with 48,185 CVEs published that year, under 0.4%.⁠6 NIST calls the shortfall expected, since CISA's scope leans toward US government systems and critical infrastructure, and warns that KEV lists may not be comprehensive.⁠11

Measured In the EPSS team's December 2022 test, a KEV-only strategy acted on 0.5% of CVEs at 53.2% efficiency and covered 5.9% of CVEs with exploitation activity, because almost 8,000 saw activity that month.⁠10 Opinion Part of that gap is definition. EPSS counts sensor-seen attempts, and KEV excludes scanning and anything short of reliable evidence by rule.

Some well-known cases sit outside the list. Socket, which sells package-malware detection, notes that the xz backdoor (CVE-2024-3094) and the March 2026 axios compromise are not in KEV. It adds that six open source compromises that did reach the list were added a median of 11 days after the first malicious release.⁠6

The week CVE-2026-111497 landed

Node.js shipped fixes for CVE-2026-111497 on 21 October. CISA added the CVE to KEV on 27 October, with a federal due date of 10 November. On 28 October, Corgea recorded an EPSS reading of 55.48%, the 99.03rd percentile.⁠8 That is six days from patch to KEV. I found no EPSS reading from before the listing, so I can't say whether the forecast led or followed the record. For a team running Node.js, the KEV flag settled whether to act. Exposure settled how fast, and the test is whether any internet-facing service decompresses Brotli data it didn't produce, in request bodies or in responses it fetches.⁠8

The queue fails before the scoring does

Reported In Verizon's 2026 DBIR, exploitation of vulnerabilities was the most common initial access vector, at 31% of breaches. Organizations in the dataset fully remediated only 26% of KEV vulnerabilities in 2025, down from 38%, and the median time to full remediation rose from 32 to 43 days.⁠13 Verizon's advice is to prioritize by active exploitation. It found that the chance of renewed exploitation falls after about 30 days, again at 90 days, and again near nine months.⁠13

Opinion Teams will argue EPSS cutoffs for a quarter while exploited CVEs age in the queue. Fix the KEV latency first. The threshold debate pays off after that.

The order to use

Opinion KEV first. Match your findings to the list, then set the clock from exposure, automation and impact. If you write your own SLA, copy the 3-day, 14-day, 60-day and next-upgrade tiers rather than inventing new ones.⁠4

EPSS sorts everything else, inside the tiers exposure gives you. Set the cutoff from capacity. The EPSS paper cites 2022 Kenna and Cyentia research putting the median monthly remediation rate at 15.5% of open vulnerabilities,⁠10 so a cutoff that yields more than your team closes in a month only builds a backlog.

CVSS comes last. Use it for impact and tie-breaks, and rescore with threat and environment metrics where you can automate them.⁠2

Exposure runs across all three, and only the asset owner knows it.⁠6

How you'd notice

The problem is already happening if your SLA table reads CVSS bands and has no KEV column. It is happening if a scanner export holds KEV matches older than 14 days. Look for CVEs that arrive with no NIST score or product list. NIST now leaves most CVEs outside its priority groups unenriched, so a scanner that returns no match for them may be missing data rather than reporting a clean result.⁠3 And if your EPSS history shows a step on 15 June, that is the model change.⁠7

What to change first

Do today

  1. Join the KEV JSON to your open findings by CVE ID. CISA publishes the catalog as JSON and CSV, and a GitHub mirror carries every change.⁠6
  2. Mark each KEV match exposed or not exposed, by asking whether an untrusted party on a public network can reach the asset.⁠4
  3. Patch or mitigate exposed KEV matches first. Check each one that also gives full control of the system for prior compromise before you close it.⁠4

Do this week

  1. Rewrite every SLA keyed to CVSS bands. Use the 3-day, 14-day, 60-day and next-upgrade tiers as starting values.⁠4
  2. Pull EPSS daily and store the model date with each reading, so a version change shows up in your history instead of your trend.
  3. List every report or integration that reads NVD CVSS or CPE data, and relabel CVEs with no NIST data as unscored, not low.⁠3
  4. Set your EPSS cutoff from what your team closes in a month, and review it each quarter.⁠10

Can wait

  1. Rescore the full inventory in CVSS v4.0 with threat and environment metrics. FIRST says that gets closer to risk, and it pays only where you can automate asset and threat data.⁠2
  2. Build NIST's Likely Exploited Vulnerabilities metric. NIST says its margin of error is unknown and it is seeking industry data to test it.⁠11
  3. Track the federal start date. Socket, citing FedRAMP's notice, puts agency remediation on the BOD 26-04 timelines at 7 December 2026, and your own clock can start earlier.⁠6