How matching works¶
This page follows one device through the engine, from the text in your inventory to a ranked fix plan. The same engine serves the web app, the API, the SDK, the MCP servers and the partner widget.
1. Resolution: from inventory text to a catalog product¶
Inventories describe devices loosely: "AB ControlLogix 5580 fw 37.011", "Siemens S7-1516-3 PN/DP", "Modicon M340 BMXP342020 2.20". CVE data names products formally, through CPE names such as cpe:2.3:h:schneider-electric:modicon_m340. The resolution layer connects the two.
It works in this order:
- Vendor. Known aliases and spellings ("AB" and "Allen-Bradley" are Rockwell Automation), exact names, normalized names, then fuzzy matching. Vendor twins (the same company under two spellings) are rolled up to one canonical vendor.
- Product. Exact name, known CPE spelling, prefix, substring, shared words; curated order-code patterns for common OT vendors (a Siemens
6ES7...order number or a Rockwell catalog number identifies the product); the vendor's part number when you send one separately. - Score. Each candidate gets a confidence from 0 to 100 and a band: HIGH (80 or more), MEDIUM (60 to 79), LOW (under 60).
- Decision. Resolved (one clear match), ambiguous (several products within 12 points of the top; the candidates are listed and no CVEs are returned until you choose) or unresolved (nothing matched well enough).
The rule throughout is precision over coverage: a retry with a looser reading is adopted only when it resolves cleanly. A device that cannot be identified is marked needs review with the reason; it never quietly becomes a device with no findings.
When only a product family is identified (for example "S7-1500" without the CPU model), the result is answered at family level, its confidence is held below HIGH, and the CVEs that apply only to some models are marked so you know to send the model.
2. Version comparison against affected ranges¶
Each CVE affects products through rows such as "version 1.0 up to and including 3.11" or "every version before 2.7". The engine compares your version with each row:
| Match tier | Meaning |
|---|---|
EXACT | Your version equals a listed affected version |
RANGE | Your version falls inside an affected range |
PRODUCT_WIDE | The record gives no version limit, or you sent no version. Not confirmed for your version |
When several rows match one CVE, the most specific tier wins.
Version schemes differ by vendor, and the engine handles them explicitly rather than comparing strings:
- Leading letters and release prefixes:
V3.5,R510,SV3.70. - Schneider Electric's two-digit minors: firmware 3.4 is read as 3.40, so it sorts below 3.5 (3.50) the way Schneider numbers its releases. Reading it as 3.4 versus 3.5 the usual way would give wrong answers at the boundaries.
- Suffixes:
1.265bor5.0(3)are compared approximately and flagged (VERSION_SUFFIX_APPROXIMATE); a space-separated suffix such asSP2is dropped and flagged. - Date-style bounds against a non-date version are never compared; the result says so.
- Windows hosts are decided by Microsoft's own fixed build per CVE. See Windows patch level.
Where a vendor's security advisory states a precise fixed release, that statement is stored with its provenance and outranks a looser "all versions" listing in a CVE record.
3. Applicability: how sure is "affected"?¶
Every CVE listed for a device carries an applicability class:
| Class | When |
|---|---|
CONFIRMED_AFFECTED | Your version was compared with the record's range or listed version, it is inside, and the model is known |
POSSIBLY_AFFECTED | Listed, but not compared: the record has no version limit, the build is not recorded, only the family was identified, or the vendor lists the CVE for some models only. Each comes with the reason |
LIKELY_NOT_AFFECTED | The record has no version limit, but the vendor's stated fixed version is at or below your version. This is an inference from the vendor's fix, not a vendor statement that you are not affected. These sort last and are flagged |
4. Fixes: what to install, and how sure we are¶
Each CVE gets a fix, chosen in a fixed order:
- The fixed version the vendor's advisory states for this product.
- A vendor statement that this product will not be fixed (mitigation only), quoted.
- A fixed version stored for the CVE.
- A version derived from the end of the affected range ("upgrade to 2.7 or later" for an exclusive end; "upgrade beyond 3.11" for an inclusive one). Derived fixes are always labelled: confirm the release in the vendor advisory, because that exact release may not exist in the product line.
- "Vendor fix available (version not recorded)".
- "No fix known".
Each fix also carries a class: VENDOR_CONFIRMED, RANGE_DERIVED, WORKAROUND_ONLY, NO_FIX_PUBLISHED, NOT_NEEDED (for likely not affected CVEs) or UNKNOWN.
CVEs with the same fix form a fix group. The fix plan answers the practical question in two lines: the one upgrade that moves past every known-exploited CVE, and the one that moves past every CVE with a version fix.
5. Priority order¶
Within a device, CVEs are ordered:
- Confirmed for your version before unconfirmed.
- Known-exploited first (listed in the known-exploited vulnerabilities catalog).
- Public exploit code available (an exploit database entry or a public repository that claims one).
- A fix available before no fix known.
- Severity score (CVSS), then exploit probability (EPSS), highest first.
Across devices, Fix first groups the same action on the same product and orders the actions: those clearing a known-exploited CVE past its catalog due date first, then those clearing more known-exploited CVEs, then by highest severity and the number of findings cleared. Your own notes (owner, criticality, tags) never change a result or its order.
6. Hashes: knowing when a result changed¶
Every device result carries two fingerprints:
result_hashcovers the identified product, the version and every (CVE, match tier) pair, with a digest of each CVE's range source and the vendor's fix and mitigation guidance, so a new vendor fix counts as a change. It does not depend on sort, filters or paging, and no link is hashed.status_hashcovers the status facts: which CVEs are known-exploited, the lifecycle answer and firmware currency.
They are SHA-256 digests of a canonical form of the result. Equal input gives equal hashes in every channel. Keep them, and use the cheap change check (a tenth of the cost of a full check) to learn which devices changed. A status can change without the findings changing, for example when a CVE you already have becomes known-exploited, or the vendor publishes an end of support date.
Measured on 10,000 devices: 50 of 50 randomly chosen stored results were identical to a fresh engine run.
7. What the engine reports when it cannot answer¶
| Situation | What you see |
|---|---|
| Product not identified | unresolved, no CVEs, needs review, the closest catalog products |
| Several possible products | ambiguous, the candidates, no CVEs until you choose |
| No version sent | Every CVE PRODUCT_WIDE, warning NO_VERSION_SUPPLIED |
| Product has no CVE data behind it | Coverage no_cpe_data: undetermined, needs review |
| Only product-wide data exists | Coverage partial: CVEs apply but versions cannot be checked |
| Version outside every known range | NO_MATCHING_RANGE, which is determined from structured sources only |
Continue with Zero CVEs is not the same as no risk.