How the mlab vuln index works

A score from 0 to 100 for how bad a CVE is in practice, not on paper. Every input is public and the rules below are the code.

Five parts

PartPointsWhat gives them
Exploitation40All of it when the CVE is in the CISA KEV or the EU KEV catalog, when CISA's SSVC says exploitation is active, or when two or more articles report it exploited. Otherwise the best of: one article reporting exploitation (30), a public proof of concept (25), or 35 × √EPSS.
Exposure25Read from the CVSS vector. Attack vector: network 10, adjacent 6, local 2, physical 0. Privileges: none 8, low 4, high 1. User interaction: none 7, required 3. No vector: 12.
Severity15The CVSS base score × 1.5. NVD's own score first, then the CNA's; 3.1, 3.0, then 4.0.
Attention15Distinct outlets that covered the CVE, from the security press read by news.mlab.sh: 3 for the first, 2 for each next one. For CVEs older than that press history, how widely attackers go after it counts instead: from an EPSS of 50%, 15 × EPSS. The higher of the two is kept.
Momentum5Added to a KEV catalog in the last 30 days, or its EPSS jumped in the last 14 days.

The boards

One score per CVE, whatever the window. The window only decides who competes: a CVE is on the 7-day, 30-day or 1-year board when, inside that window, it entered a KEV catalog, its EPSS jumped, the press covered it, or it was published and is critical or likely to be exploited (EPSS percentile 90 for 7 and 30 days, 95 for a year). The all-time board holds every KEV entry and the top 1% by EPSS.

The 50 highest scores make the board. Ties go to the higher EPSS, then the newer CVE. Boards are rebuilt every hour; the last board of each ISO week is archived, and the arrows compare with the week before.

Labels

Every CVE with a score also gets two labels, from fixed rules. The rules are versioned (version 1) and never set by hand.

LabelWhen
Actively exploitedIn CISA KEV or the EU KEV, exploitation active per CISA SSVC, or two or more articles reporting exploitation.
Default setupsReachable over the network with no login, no user action and no special configuration: CVSS attack vector network, complexity low, no privileges, no user interaction (and, in CVSS 4.0, no attack requirements).
Specific setups onlyIt needs one of those: a non-default setting, a login, a user to do something, or local access.
Exposure unclearNo CVSS vector to read yet.
Needs attentionAny of: added to CISA KEV (or the EU KEV) in the last 30 days; a CISA remediation deadline still ahead; EPSS jumped in the last 14 days and is now 10% or more; the first press report of an exploit or proof of concept in the last 14 days; an index of 80 or more for a CVE published in the last 30 days.
No labelNone of the above: a record to look things up in.

Next to the risk score

Each CVE page keeps its risk score (CVSS × EPSS percentile, 100 once exploited), which answers "should this be patched". The index answers "how much does it matter right now": it spreads apart the thousands of exploited CVEs the risk score pins at 100, and it rewards what the attackers and the press are actually doing.

Limits

  • Exposure reads the vector as published. It cannot know whether the affected setup is common.
  • Attention counts outlets news.mlab.sh reads; a CVE covered elsewhere only is under-scored.
  • A CVE NVD has not received yet has no score until it does.