On this page
A vulnerability scan of a modest estate can return hundreds of findings, and most teams sort them the same way: by CVSS score, highest first. It feels objective. It is also the reason the Outlook flaw that attackers were using last week can sit below a CVSS 9.8 on a server nobody can reach.
This post is for whoever owns the patching backlog: IT managers, security leads and the people who write the vulnerability management policy. It explains what CVSS, EPSS and CISA's Known Exploited Vulnerabilities (KEV) catalogue each measure, why none of them is enough alone, and how we combine them with business context into a short set of rules a team can follow. To keep it concrete, we ran a small triage script over a sample scanner export using the real KEV and EPSS data published on 13 November 2026.
Why sorting by CVSS fails
CVSS (the Common Vulnerability Scoring System, maintained by FIRST) gives each vulnerability a score from 0 to 10. The score most tools show is the Base score. It describes the flaw in a vacuum: how it is reached, how complex the attack is, and what it could do to confidentiality, integrity and availability.
That is useful, and it is honest about its limits. The CVSS v4.0 specification describes its outputs as scores "indicating the severity of a vulnerability relative to other vulnerabilities". It also says consumers "should enrich the Base metrics with Threat and Environmental metric values". In other words, CVSS Base is a severity rating, not a risk rating. It does not know whether anyone is exploiting the flaw, and it does not know where the affected software sits in your network.
The practical problem is volume. So many vulnerabilities score 7.0 or above that "fix everything High and Critical within 30 days" quietly becomes "fix nothing on time".
Three signals that answer different questions
EPSS: how likely is exploitation soon?
The Exploit Prediction Scoring System, also run by FIRST, publishes a daily score for every CVE. In FIRST's words, EPSS "estimates the probability a vulnerability will be exploited in the wild within the next 30 days". The score runs from 0 to 1, and FIRST also publishes a percentile, which shows how the score compares with every other CVE.
The EPSS FAQ is direct about what EPSS is not. It covers likelihood only and is "not a complete risk score". It also warns against multiplying EPSS by CVSS, because CVSS is an ordinal ranking rather than a calibrated measure, so the product means nothing.
KEV: has exploitation already been confirmed?
CISA's Known Exploited Vulnerabilities catalogue is a list, not a score. A flaw is added only when it has a CVE ID, there is "reliable evidence" of active exploitation in the wild, and there is a clear fix. US federal agencies must remediate entries by a due date. CISA recommends that every other organisation build prompt KEV remediation into its own vulnerability management plan.
KEV is narrow on purpose. In the catalogue we downloaded on 14 November 2026 there were 1,781 entries, against 406,184 CVEs carrying an EPSS score. That is under half of one per cent. The same file also marks 361 entries as known to be used in ransomware campaigns, a useful extra flag.
The EPSS FAQ itself says a KEV listing should outweigh a low EPSS score for the same CVE. Our data shows why that matters:
$ python3 stats.py kev.json epss_scores-current.csv.gz#model_version:v2026.06.15,score_date:2026-11-13T12:03:22ZCVEs with an EPSS score: 406,184 EPSS >= 0.10: 17,555 (4.3%)CVEs in KEV: 1,781 (0.44% of scored CVEs) KEV CVEs with EPSS < 0.10: 441 (24.8%)EPSS >= 0.10 CVEs that are in KEV: 1,340 (7.6%)A quarter of confirmed-exploited vulnerabilities score below 10% on EPSS. Most CVEs with a high EPSS score are not in KEV. The two signals overlap, but far from completely. A policy that relies on either one alone will miss things the other catches.
Business context: does this asset matter, and can attackers reach it?
The third input is yours, and no feed can supply it. Is the system on the internet, on user devices that open email and web pages, or on an internal segment? Does it hold customer data, run payments or control identity? The same flaw matters more on the VPN gateway than on a build container.
The three side by side
| CVSS Base | EPSS | CISA KEV | |
|---|---|---|---|
| Question it answers | How bad could this be? | How likely is exploitation in the next 30 days? | Has exploitation been confirmed? |
| Form | Score 0–10 | Probability 0–1, plus percentile | Yes or no, with a due date |
| Updated | Rarely, once scored | Daily | When CISA adds entries (often several times a week) |
| Main blind spot | Ignores real-world exploitation and your environment | Says nothing about impact or your environment | Only covers confirmed cases with a known fix; lags new campaigns |
| Best use | Breaking ties; judging impact | Ordering the long tail | Moving items to the front of the queue |
A worked example from our triage script
To test the rules before recommending them, we built a sample scanner export of twelve findings across invented assets. The CVEs are real. The CVSS scores come from the NVD, and the EPSS and KEV data come from the same-day downloads above. Each asset has two pieces of context: exposure (internet, user endpoint or internal) and criticality (high, medium or low).
The heart of the script is a short, readable rule set. This is the logic we suggest writing into a policy, almost word for word:
def tier(f):
in_kev = f["cve"] in kev
reachable = f["exposure"] in ("internet", "user-endpoint")
likely = f["epss"] >= 0.10 # >=10% chance of exploitation activity in 30 days
if in_kev and reachable:
t = 1 # known exploited and attackers can reach it: days
elif in_kev or (likely and reachable):
t = 2 # strong signal, but one condition missing: this sprint
elif f["cvss"] >= 7.0 or likely:
t = 3 # severe on paper, no exploitation evidence: normal cycle
else:
t = 4
if f["criticality"] == "high" and t > 2:
t -= 1 # business context moves a crown-jewel asset up one tier,
# but never into P1: P1 needs confirmed exploitation
return t
Our first version let business context lift any finding by one tier, including into P1. That put Terrapin (CVE-2023-48795) on the main website into P1. It has an EPSS score of 0.935, but a successful attack needs someone sitting in the network path between client and server. A high EPSS score says exploitation activity is likely. It does not say a successful attempt is damaging. So we changed the rule: P1 requires confirmed exploitation, and context can only move findings within the lower tiers.
Here is what the script produced:
$ python3 triage.py kev.json epss_scores-current.csv.gzSorted by CVSS (the usual report order): 1. CVE-2024-3094 CVSS 10.0 dev-vm-07.example.com 2. CVE-2022-22965 CVSS 9.8 api.example.com 3. CVE-2024-21762 CVSS 9.8 vpn.example.com 4. CVE-2023-38545 CVSS 9.8 etl-02.example.com 5. CVE-2024-38063 CVSS 9.8 fileserver-01.example.com 6. CVE-2026-125350 CVSS 8.8 browser-fleet 7. CVE-2026-123158 CVSS 8.1 build-runner-01.example.com 8. CVE-2026-119977 CVSS 8.1 blog.example.com 9. CVE-2024-6387 CVSS 8.1 jump-01.example.com 10. CVE-2023-44487 CVSS 7.5 www.example.com 11. CVE-2026-124194 CVSS 6.5 mail-client-fleet 12. CVE-2023-48795 CVSS 5.9 www.example.com Sorted by tier, then EPSS: tier cve cvss epss pctl kev exposure crit asset P1 CVE-2023-44487 7.5 1.000 1.000 yes internet high www.example.com P1 CVE-2022-22965 9.8 0.996 0.999 yes internet high api.example.com P1 CVE-2024-21762 9.8 0.846 0.997 yes internet high vpn.example.com P1 CVE-2026-124194 6.5 0.016 0.765 yes user-endpoint high mail-client-fleet P1 CVE-2026-119977 8.1 0.012 0.686 yes internet low blog.example.com P1 CVE-2026-125350 8.8 0.010 0.622 yes user-endpoint medium browser-fleet P2 CVE-2024-6387 8.1 0.995 0.999 no internal high jump-01.example.com P2 CVE-2023-48795 5.9 0.935 0.998 no internet high www.example.com P2 CVE-2024-38063 9.8 0.736 0.995 no internal high fileserver-01.example.com P2 CVE-2026-123158 8.1 0.189 0.973 yes internal medium build-runner-01.example.com P3 CVE-2024-3094 10.0 0.909 0.998 no internal low dev-vm-07.example.com P3 CVE-2023-38545 9.8 0.783 0.996 no internal low etl-02.example.comThree things stand out.
The Outlook flaw moves from eleventh to the top tier. CVE-2026-124194 has a CVSS of 6.5 and an EPSS of only 0.016. It is also in KEV and sits on every user's laptop. That is the flaw we covered in our Outlook zero-day threat brief. A CVSS-sorted report would have put it near the bottom. EPSS alone would have put it there too.
Three of the six P1 findings have EPSS scores below 0.02. The Outlook, WordPress and Chrome flaws were all added to KEV in the past fortnight. EPSS is a model trained on past activity, so it can take time to catch up with a new campaign. KEV records the confirmed exploitation directly. This is why the rule order matters: check KEV first.
The two top CVSS scores land in P3. The xz-utils backdoor (CVSS 10.0) and the libcurl flaw (9.8) sit on low-value internal machines and are not in KEV. They still get fixed, on the normal cycle. A tier is the deadline, not permission to ignore something. If a P3 fix is a single package update, we tell teams to just do it.
What a good prioritisation policy looks like
The script is only a way of testing rules. What lasts is the policy. When we review vulnerability management as part of a governance, risk and compliance engagement, these are the elements we look for:
- Tiers with deadlines, decided in advance. For example: P1 within 72 hours or the KEV due date, whichever is sooner; P2 within 14 days; P3 within 30 days; P4 within 90 days. The numbers are yours to set. What matters is that they are written down and agreed by the business. KEV due dates for this month's additions ranged from 3 to 14 days, which gives a sense of the pace for P1.
- An asset inventory with two context fields. Exposure and criticality per asset are enough to start. Without them every rule falls back to CVSS. If you do not know what you expose, our post on security basics that stop most attacks covers where to start.
- Feeds pulled daily, not quarterly. EPSS updates every day and KEV changes several times a week. A score captured at scan time goes stale within days. Both feeds are free and machine-readable, as the script above shows.
- A documented exception route. Some fixes cannot be applied on time. The policy should say who can accept the risk, what compensating control is required (blocking a port, disabling a feature, extra monitoring) and when the exception expires.
- Metrics on the tiers, not on raw counts. "P1 items closed within deadline" tells leadership far more than "1,200 open vulnerabilities". Report the share of each tier closed on time and the age of the oldest open P1.
- A periodic review of the rules. As our first version showed, rules can produce odd results. Review a sample of tier decisions each quarter.
If you follow ISO 27001, this maps onto the technical vulnerability management control (Annex A 8.8 in the 2022 edition), and a written tiering rule with evidence of on-time closure is the evidence auditors look for. Larger teams may prefer CISA's SSVC (Stakeholder-Specific Vulnerability Categorization), which uses decision trees for the same purpose.
Questions to ask your team this week
- When a new KEV entry matches our software, how long until someone knows, and who is that person?
- Can we list which assets are internet-facing and which hold sensitive data?
- Does our scanner or ticketing system show EPSS and KEV status next to CVSS? If not, can it?
- What were our three oldest open findings last month, and why are they still open?
- Do our patching deadlines appear in a policy that leadership signed, or only in someone's head?
None of this needs a new platform. It needs a short rule set, two free data feeds and an inventory that knows what matters. When we test this in practice, the inventory is almost always the hardest part.
Sources
- FIRST, CVSS v4.0 Specification Document
- FIRST, EPSS FAQ and the daily EPSS data file (scores dated 13 November 2026)
- CISA, Known Exploited Vulnerabilities Catalog (JSON feed, catalogue version 2026.11.13)
- CISA, Stakeholder-Specific Vulnerability Categorization (SSVC)
- NIST, National Vulnerability Database (CVSS base scores, fetched through its CVE API)
Written by
Haris Siddique
GRC Consultant, FutaqSec