diff --git a/app/services/app_cve_scanner_service.py b/app/services/app_cve_scanner_service.py
index c8c7e2c..55b5120 100644
--- a/app/services/app_cve_scanner_service.py
+++ b/app/services/app_cve_scanner_service.py
@@ -119,14 +119,14 @@ _REGISTRY: List[tuple] = [
 # version (13/14/15) without ranges, so it needs the Intune security-patch
 # level + Android bulletin parsing — a separate feature.
 # ponytail: iOS/iPadOS only; add Android when the patch-level path is built.
-# NOTE: Microsoft OS/products are deliberately NOT in this registry. NVD models
-# them as rangeless CPEs (`cpe:2.3:o:microsoft:windows_server_2016:-:*` with no
-# versionEndExcluding) and cvelistV5 MS records use `lessThan: "publication"` —
-# neither carries the fixed BUILD, so neither can tell a patched host from an
-# unpatched one. _in_range() correctly rejects those wildcard/`-` versions, so
-# adding them here would match nothing while burning NVD quota. MSRC (KB +
-# FixedBuild per product) is the only source with MS patch state — see the
-# msrc remediation path.
+# Windows OS CVEs are handled by cvelistv5_scan_service.scan_asset_os(), not
+# from here. Modern Microsoft CVE records DO carry real build ranges (e.g.
+# CVE-2026-47291: 20/20 affected entries with a numeric lessThan) — but NVD
+# flattens them to an END bound only (versionEndExcluding, no start), so a
+# 1607 host (14393.x) would fall inside the 22H2 range (endExcluding
+# 19045.7417) and false-positive. cvelistV5 keeps the range BOUNDED
+# (version 10.0.22631.0 .. lessThan 10.0.22631.7219), and those bounds pick the
+# host's release on their own — which is why the OS path lives there.
 _OS_REGISTRY: List[tuple] = [
     (re.compile(r"ipad", re.I), {"key": "cpe:apple:ipados",
                                  "cpe": "cpe:2.3:o:apple:ipados", "label": "Apple iPadOS"}),
@@ -647,6 +647,19 @@ def run_app_cve_scan(db: Session, asset_id: Optional[int] = None) -> dict:
         except Exception as e:
             stats["errors"].append(f"asset {asset.id} os: {e}")
 
+        # Windows OS CVEs from cvelistV5's bounded build ranges — no software
+        # inventory needed, so this runs for every Windows asset.
+        if cve5_index:
+            try:
+                from app.services import cvelistv5_scan_service
+                n = cvelistv5_scan_service.scan_asset_os(
+                    db, asset, cve5_index, new_ids, touched=touched_cves)
+                if n:
+                    stats["findings"] += n
+                    touched = True
+            except Exception as e:
+                stats["errors"].append(f"asset {asset.id} win-os: {e}")
+
         # Package-level CVEs (Wazuh syscollector / Intune detectedApps).
         packages: list = []
         try:
diff --git a/app/services/cvelistv5_scan_service.py b/app/services/cvelistv5_scan_service.py
index f6f7b2c..0584760 100644
--- a/app/services/cvelistv5_scan_service.py
+++ b/app/services/cvelistv5_scan_service.py
@@ -42,7 +42,7 @@ logger = logging.getLogger(__name__)
 _ZIP_PATH = "/tmp/truevuln-cvelistv5-cache.zip"
 _ZIP_URL = "https://github.com/CVEProject/cvelistV5/archive/refs/heads/main.zip"
 _ZIP_TTL = 12 * 3600
-_INDEX_SETTING = "cvelistv5_product_index_v3"  # v3: entries now carry cvss/severity
+_INDEX_SETTING = "cvelistv5_product_index_v4"  # v4: pattern products (Windows OS)
 _INDEX_TTL = timedelta(hours=26)  # rebuilt nightly; a missed night still serves
 
 # Curated registry: name-regex (installed software) → cvelistV5 (vendor,
@@ -96,6 +96,38 @@ for _e in _REGISTRY:
     for _v, _p in _e["pairs"]:
         _PAIR_TO_KEY[(_v.lower(), _p.lower())] = _e["key"]
 
+# Pattern products — for families whose CVE records name one product per
+# RELEASE, so an exact pair list would need dozens of rows and rot every time
+# Microsoft ships a new one ("Windows 10 Version 1607", "Windows 11 Version
+# 23H2", "Windows Server 2016 (Server Core installation)", ...).
+#
+# Windows is the OS case: the record's ranges are BOUNDED
+# (version 10.0.22631.0 .. lessThan 10.0.22631.7219), so a host's build falls
+# into exactly the range of its own release — the bounds do the release
+# selection, no build→release table needed, and Server and Client are the same
+# problem. (NVD flattens these to an end bound only, which is why the OS scan
+# reads cvelistV5 and not NVD-CPE.)
+_PRODUCT_PATTERNS: List[dict] = [
+    {"key": "windows", "vendor_re": r"^microsoft$",
+     "product_re": r"^windows\s+(10|11|server)\b"},
+]
+_PATTERNS_COMPILED = [
+    (re.compile(p["vendor_re"], re.I), re.compile(p["product_re"], re.I), p["key"])
+    for p in _PRODUCT_PATTERNS
+]
+
+
+def _pair_key(vendor: str, product: str) -> Optional[str]:
+    """(vendor, product) from a CVE record → curated product key."""
+    v, p = (vendor or "").strip().lower(), (product or "").strip().lower()
+    key = _PAIR_TO_KEY.get((v, p))
+    if key:
+        return key
+    for vrx, prx, k in _PATTERNS_COMPILED:
+        if vrx.search(v) and prx.search(p):
+            return k
+    return None
+
 
 def resolve(name: str) -> Optional[str]:
     n = (name or "").strip().lower()
@@ -278,8 +310,7 @@ def build_product_index(db: Session) -> dict:
                 continue
             cvss, sev = _cvss_from_record(data)
             for aff in affected:
-                key = _PAIR_TO_KEY.get(((aff.get("vendor") or "").lower(),
-                                        (aff.get("product") or "").lower()))
+                key = _pair_key(aff.get("vendor") or "", aff.get("product") or "")
                 if not key:
                     continue
                 plats = [str(p).lower().strip() for p in (aff.get("platforms") or [])]
@@ -436,6 +467,45 @@ def suppress_false_positives(db: Session, asset_id: Optional[int] = None) -> dic
     return stats
 
 
+def scan_asset_os(db: Session, asset, index: dict,
+                  new_ids: Optional[list] = None, touched: Optional[set] = None) -> int:
+    """Windows OS CVEs straight from the asset's build (asset.os_version).
+
+    Works for Server AND Client without a build→release table: every affected
+    entry is a bounded range for one release (10.0.22631.0 .. 10.0.22631.7219),
+    so a host's build only ever falls inside its own release's range. Releases
+    that share a build line (Windows 10 1607 / Server 2016 = 14393; Windows 11
+    24H2 / Server 2025 = 26100) take the same fix anyway, so matching either is
+    correct.
+    """
+    if not index:
+        return 0
+    entries = index.get("windows") or []
+    if not entries:
+        return 0
+    if cpe._os_family(asset.operating_system or "") != "windows":
+        return 0
+    cver = cpe._clean_version(asset.os_version or "")
+    if not cver:
+        return 0
+    if new_ids is None:
+        new_ids = []
+    label = (asset.operating_system or "Microsoft Windows").strip()
+    count = 0
+    for entry in entries:
+        if not _affected(cver, entry.get("start"), entry.get("lt"), entry.get("lte")):
+            continue
+        c = {"cve": entry["cve"], "cvss": entry.get("cvss"), "severity": entry.get("sev"),
+             "fixed": entry.get("lt")}
+        try:
+            before = len(new_ids)
+            cpe._upsert(db, asset, label, asset.os_version or cver, c, new_ids, touched=touched)
+            count += 1 if len(new_ids) > before else 0
+        except Exception as e:
+            logger.debug("cvelistv5 OS upsert failed (%s on %s): %s", entry["cve"], asset.id, e)
+    return count
+
+
 def scan_asset(db: Session, asset, packages: list, index: dict,
                new_ids: Optional[list] = None, touched: Optional[set] = None) -> int:
     """Match an asset's installed software against the cvelistV5 index.