brace-expansion: Quadratic-time expansion of the `{a},b}` rewrite causes CPU denial of service

Summary

Expanding {a},b}-shaped input takes time quadratic in the number of literal } characters, blocking the event loop.

Bash preserves a quirk where a brace group followed by a comma set still expands ({a},b}). The parser implements this by rewriting the string and restarting the scan. Each pass absorbs exactly one } and re-scans from the beginning, so n trailing braces cost n full passes.

Reproduction

const build = n => '{a}' + '}'.repeat(n) + ',z}'

for (const n of [8000, 16000, 32000, 64000, 128000]) {
  const t = Date.now()
  expand(build(n))
  console.log(n, Date.now() - t + 'ms')
}

n	input	time	results
8,000	8 KB	110 ms	2
16,000	16 KB	446 ms	2
32,000	32 KB	1.7 s	2
64,000	64 KB	6.9 s	2
128,000	128 KB	27.7 s	2

ms/n^2 is flat at ~1.7 and each doubling of n costs exactly 4.0x - quadratic. 128 KB of input blocks the event loop for nearly half a minute to produce two results.

Mechanism

Instrumenting the rewrite branch confirms it runs exactly n + 1 times, once per literal }, each re-scanning the whole string.

There is a second multiplier. The rewrite replaces the group's closing } with the internal escClose sentinel, which is '\0CLOSE' + Math.random() + '\0' - about 25 characters. The working string therefore grows by ~25 characters on every pass:

n	input length	final string length
1,000	1,006	26,006
8,000	8,006	208,006

So the input is inflated roughly 26x, and that factor multiplies both the quadratic constant and peak memory. This makes it partly a memory-pressure issue as well as a CPU one.

Why max and maxLength do not help

The cost is in parsing, before the result set exists. The payload yields 2 results regardless of size, so neither bound is ever reached.

Impact

An application passing an untrusted pattern to expand(), directly or through minimatch / glob, can have its event loop blocked for tens of seconds by a payload well under minimatch's 65,536-character cap. For a single-threaded Node server that is a full stall, not just a slow request.

Degraded availability rather than a crash - the process recovers once the expansion completes.

Affected versions

Verified affected on 1.1.18, 2.1.4, 3.0.6 and 5.0.9, all within a few percent of each other (~460-490 ms at n=16,000).

Patch

The rewrite loop gets an iteration bound. Past the cap the remaining string is treated as non-expanding and returned literally, consistent with the existing max / maxLength caps, which truncate rather than throw.

Note this bounds the number of passes, not the cost of each: worst-case work remains proportional to cap x input length. The cap is set low enough that the residual is bounded in practice, and far above what any realistic {a},b} input needs.

Severity note

Scored 5.3 Medium (A:L) for consistency with GHSA-3jxr-9vmj-r5cp, the other algorithmic-complexity advisory on this package (CWE-407), which uses the same vector. The stack-exhaustion advisories on this package score A:H because they crash the process outright; this one stalls it.