zlib: harden ZIP archive reading and writing against malicious input

[pipobscure]

Summary

This PR fixes four security issues in the node:zlib ZIP support (ZipEntry/ZipFile/ZipBuffer), found by an audit of the parser and the on-disk reader/writer. Two are high-severity (a parser-confusion divergence and a denial-of-service hang); two are medium-severity lifecycle/robustness bugs. Each issue is landed test-first: a commit adds a failing regression test, followed by the fix that turns it green.

The ZIP surface treats an archive (bytes in a Buffer, or a file on disk) and every header field as fully attacker-controlled, so these are all reachable from untrusted input.

---

1. Local vs. central header disagreement — parser confusion (high)

Issue. The reader treats the central directory as authoritative for a member's compression method, sizes, CRC, and name, but only validated the local file header's signature. It never checked that the two headers agree. Two consequences:

• Different bytes than other tools. Because Info-ZIP unzip extracts using the local header's method/size, an archive whose local and central headers disagree yields different content in Node than in unzip for the same member name — a classic inspect-then-consume bypass (a scanner sees content Q; Node consumes content P). Verified against unzip/python -m zipfile.
• Encryption-flag confusion. The "is this member encrypted?" decision was read from the local header while the member's identity came from the central one. An attacker could mark a member encrypted in the central directory (so a central-directory-based scanner such as Python's zipfile treats it as opaque and skips it) while leaving the local flag clear, so Node silently decoded the plaintext payload.

Solution. When a member is read, cross-check the local file header against the central entry — compression method, CRC-32, compressed size, and uncompressed size (with a spec-compliant exemption for data-descriptor entries, whose local CRC/sizes are legitimately zero), plus the encryption flag — and reject any disagreement with ERR_ZIP_INVALID_ARCHIVE. This follows the module's existing "reject rather than silently choose one interpretation" stance and removes Node from both sides of any local/central divergence.

The existing hardening/coverage tests that forged a decode-time size/CRC/Zip64 lie in the central header alone now trip this earlier check; they were updated to patch both headers consistently so they still exercise the decode-time guards they target.

2. zipFiles() hangs on a FIFO/special source — DoS (high)

Issue. When archiving files, zipFiles() opened each source and only then fstat-ed it to confirm a regular file. open(2) on a FIFO blocks until a writer appears, so the guard ran too late — a FIFO source (reachable via a source path an attacker influences, or a regular-file→FIFO TOCTOU) hung the call indefinitely and pinned a libuv threadpool thread; a handful stalls all fs in the process. The code comment even claimed the guard prevented this.

Solution. Open with O_NONBLOCK so the open returns promptly and the fstat/regular-file check can reject a FIFO, device, or socket before any read. O_NONBLOCK has no effect on a regular file's subsequent reads, so the happy path is unchanged.

3. ZipFile read racing close() — use-after-close / fd reuse (medium)

Issue. A ZipEntry read runs on a file descriptor shared with its ZipFile, but reads didn't participate in close()'s lifecycle. close() marked the handle closed and released the fd while a read was still in flight, so the read landed on a closed — or, once the OS reused the fd number, another file's — descriptor (surfacing as EBADF, or a cross-file read), contradicting the class's own invariant.

Solution. Track in-flight reads on the shared handle. close() now marks the handle closing (rejecting new reads immediately), waits for in-flight reads to finish on the still-open fd, and only then releases it. closeSync(), which cannot wait, refuses while an asynchronous read is outstanding.

4. Failed directory rewrite in add() leaves a corrupt archive (medium)

Issue. add()/addEntrySync() advanced the central-directory offset and adopted the new entry into memory before the final directory rewrite. If that rewrite failed (e.g. ENOSPC/EIO after the member bytes were already written), the in-memory state and the on-disk archive were left diverged and half-updated with no restore, corrupting the next add().

Solution. Wrap the rewrite: on failure, restore the previous offset and directory entry and rewrite the original directory back, leaving the archive and the handle exactly as they were before the call, then rethrow.

---

A note on streaming reads

The audit also flagged that contentIterator() applies no default size cap. On inspection this is already safe and needs no change: the streaming decoder hard-bounds output to the member's declared uncompressed size (ERR_ZIP_ENTRY_CORRUPT the moment it inflates past it), and that size is exposed as entry.size before a byte is streamed — so it's a ceiling a consumer can inspect and trust. Rather than bolt on an arbitrary global cap (which would break legitimate large streams), this PR adds a regression test locking in that declared-size guard.

Testing

• New regression tests: test-zlib-zip-security-hardening.js (header confusion, FIFO hang, streaming bound) and test-zlib-zip-file-lifecycle.js (close-vs-read, add rollback), the latter using deterministic fs.writeSync failure injection.
• The full test/parallel/test-zlib* suite passes, including the updated test-zlib-zip-hardening.js / test-zlib-zip-coverage.js.

Commits

Structured as test-then-fix per issue:

test: cover ZIP header-confusion and DoS guards
zlib: reject local/central ZIP header mismatch
zlib: do not hang archiving a FIFO or device
test: cover ZIP fd lifecycle and add rollback
zlib: let a ZipFile read finish before close()
zlib: roll back a failed add() directory rewrite

[mcollina]

lgtm

[nodejs-github-bot]

CI: https://ci.nodejs.org/job/node-test-pull-request/75467/

[codecov]

Codecov Report

❌ Patch coverage is 83.65385% with 34 lines in your changes missing coverage. Please review.
✅ Project coverage is 90.28%. Comparing base (31cde9f) to head (719c83c).
⚠️ Report is 37 commits behind head on main.

Files with missing lines	Patch %	Lines
lib/internal/zip/file.js	58.90%	30 Missing ⚠️
lib/internal/zip/entry.js	94.28%	4 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main   #65016      +/-   ##
==========================================
- Coverage   90.30%   90.28%   -0.02%     
==========================================
  Files         759      759              
  Lines      247621   247799     +178     
  Branches    46672    46727      +55     
==========================================
+ Hits       223603   223721     +118     
- Misses      15473    15551      +78     
+ Partials     8545     8527      -18     

Files with missing lines	Coverage Δ
lib/internal/zip/archive.js	99.69% <100.00%> (+0.31%)	⬆️
lib/internal/zip/headers.js	97.52% <100.00%> (+0.26%)	⬆️
lib/internal/zip/entry.js	98.11% <94.28%> (-0.33%)	⬇️
lib/internal/zip/file.js	93.09% <58.90%> (-3.32%)	⬇️

... and 44 files with indirect coverage changes

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.

[pipobscure]

This goes together with #65002 and #65007 which fix similar issues. The three should be merged soonish so as to preclude unhardened release.

[nodejs-github-bot]

CI: https://ci.nodejs.org/job/node-test-pull-request/75555/

[pipobscure]

The CI seems stuck. And I can’t look at what failed. If there is something I can do, please ping me!