Is brace-expansion safe?

GHSA-6j4f-fj2g-mc7p

parseCommaParts() can exhaust the native stack and crash the process. There are two distinct ways to trigger it, both reachable from a single untrusted pattern string.

This is the parsing-side counterpart to CVE-2026-14257 / GHSA-mh99-v99m-4gvg. That fix made expand_() iterative and documented a constant-stack-depth guarantee, but parseCommaParts() was left recursive, so the guarantee only held for one of the two parsing paths.

parseCommaParts() recursed on the remainder of the string once per brace group:

const postParts = parseCommaParts(post)   // unbounded

A brace group containing many comma-separated groups drives one recursion level per group:

expand('{' + '{a},'.repeat(7000) + 'b}')
// RangeError: Maximum call stack size exceeded

About 7,300 repetitions - roughly 29 KB of input - is enough on Node 24; roughly 6,300 (25 KB) on Node 18. The threshold is identical on every affected release line.

Even with the recursion removed, parseCommaParts() spread whole arrays into an argument list:

p.push.apply(p, postParts)
parts.push.apply(parts, p)

Function.prototype.apply places one argument per element on the stack, so a single large array overflows it. This needs no recursion depth at all - the following reaches a recursion depth of exactly 1:

expand('{{x},' + 'a,'.repeat(125000) + 'b}')
// RangeError: Maximum call stack size exceeded

Threshold is about 124,300 repetitions (~249 KB). This vector was not part of the original report; it was found while verifying the fix. A patch that only de-recurses but keeps push.apply leaves a working denial of service behind.

Both crashes happen during parsing, before any expansion. The payloads produce one result per group, so output size grows linearly with input and is never the limiter. expand(payload, { max: 1, maxLength: 1 }) still overflows.

Any application that passes an untrusted string to expand() - directly, or through minimatch / glob where it is a user-supplied glob pattern - can be crashed. In Node, a RangeError that the application does not catch terminates the process, so a server that globs user input is exposed to remote unauthenticated denial of service.

minimatch's own MAX_PATTERN_LENGTH cap (65,536) does not help against vector 1: the overflow threshold sits well below it. Confirmed on minimatch 10.2.6 - a 64,003-byte pattern passes the length check and overflows both minimatch.braceExpand() and new minimatch.Minimatch().

This is an availability-only issue. No code execution and no data exposure.

5.0.8 and 5.0.9 overflow at the same repetition count, so the gap predates the recent advisories; those fixes simply did not reach it. Verified affected on 1.1.18, 2.1.4, 3.0.6, 5.0.8 and 5.0.9, all at an identical threshold.

parseCommaParts() is rewritten as a loop that carries the partial part across chunks, and every array append uses an element-by-element loop rather than push.apply. The redundant if (!str) return [''] guard is dropped - the loop returns [''] for the empty string on its own.

Equivalence of the old and new implementations was checked by differential testing: exhaustive over every string of {, }, ,, a up to length 7 plus 300,000 random inputs - 322,000 cases, zero mismatches.

Scored 7.5 High under CVSS 3.1 for consistency with the other availability advisories on this package (GHSA-mh99-v99m-4gvg, GHSA-rgw5-rvv9-x895), which use the same vector. The reporter self-assessed 6.9 Medium under CVSS 4.0 (CVSS:4.0/AV:N/AC:L/AT:N/PR:N/UI:N/VC:N/VI:N/VA:L/SC:N/SI:N/SA:N).

Reported by baeseungwon1010, with a working proof of concept and a proposed patch. Vector 2 was identified during maintainer verification.