transpiler cache: always verify output/sourcemap/esm_record hashes on load by robobun · Pull Request #35102 · oven-sh/bun

robobun · 2026-07-22T08:59:33Z

What

Entry::load() compared the cached output against metadata.output_hash only when the stored value was non-zero:

if self.metadata.output_hash != 0 && hash(bytes) != self.metadata.output_hash {
    return Err(crate::CrateError::InvalidHash);
}

The same != 0 guard applied to esm_record_hash, and sourcemap was never verified at all.

Why it matters

A .pile entry is addressed by hex(wyhash(source)), and its binding to the input is input_hash + input_byte_length + features_hash, all of which are derivable from the public source bytes and a copy of Bun. The only check on the transpiled output itself is output_hash, and a stored zero disabled it.

Anyone who can write to the transpiler cache directory can therefore take any known source file (for example a widely-used node_modules module), compute its cache path, prepend a payload to the cached output, zero output_hash, and have that payload execute the next time the genuine source is run.

Reproduction (lands in test/cli/run/transpiler-cache.test.ts):

run bun once on a >=4 KiB file -> .pile written
rewrite the .pile: prepend console.log("PLANTED") to the output section,
  shift sourcemap/esm_record offsets, write output_hash = 0
run bun again on the untouched source file

Before: stdout is PLANTED\nclean.
After: the entry fails InvalidHash, is unlinked, the source is re-transpiled, stdout is clean.

Fix

Entry::save() has always written output_hash, sourcemap_hash, and (for non-empty records) esm_record_hash, so the loader can compare unconditionally:

drop the != 0 guard on output_hash in all three encoding branches
drop the != 0 guard on esm_record_hash
verify sourcemap_hash when sourcemap_byte_length > 0
in the Latin-1 branch, check the short-read before hashing so the hash never sees uninitialised bytes

No format change and no version bump: every entry a released Bun has written already carries the values the loader now compares against.

The existing out-of-range-esm-record-index test previously relied on zeroing esm_record_hash to reach the deserializer; it now recomputes the hash with Bun.hash.wyhash(bytes, 42n) so the corrupted record still gets that far.

Scope

This removes the zero-hash bypass. A writer who controls the cache directory can still forge a matching wyhash, so a per-user keyed hash and an ownership/mode check on the @t@ directory (as bunx does for its temp cache) remain as follow-up hardening. Related: #32741 (this addresses the zero-hash shortcut; the input_hash collision in PoC6 and the keyed-hash hardening are separate).

coderabbitai · 2026-07-22T08:59:39Z

Walkthrough

Changes

The transpiler cache now always validates hashes for cached output code, sourcemaps, and ES module records. Tests cover zeroed output hashes and corrupted module-record indices.

Transpiler cache integrity

Layer / File(s)	Summary
Unconditional cache byte validation `src/jsc/RuntimeTranspilerCache.rs`	UTF-8, Latin-1, and UTF-16 output, sourcemaps, and ES module records now raise `InvalidHash` whenever their computed hashes differ from metadata.
Corruption defense coverage `test/cli/run/transpiler-cache.test.ts`	Tests mutate cached metadata to zero `output_hash`, verify re-transpilation without executing injected payloads, and preserve valid hashes while testing corrupted module-record indices.

Suggested reviewers: jarred-sumner

🚥 Pre-merge checks | ✅ 4

✅ Passed checks (4 passed)

Check name	Status	Explanation
Linked Issues check	✅ Passed	The cache-integrity fix and regression test align with the linked issue's reported startup crash path by rejecting tampered cache entries.
Out of Scope Changes check	✅ Passed	The changes stay focused on transpiler-cache integrity checks and related tests, with no clear unrelated additions.
Title check	✅ Passed	The title clearly summarizes the main change: unconditional transpiler cache hash verification on load.
Description check	✅ Passed	The description is detailed and covers what, why, fix, scope, and reproduction, though it lacks an explicit verification section.

_{Comment @coderabbitai help to get the list of available commands.}

robobun · 2026-07-22T08:59:39Z

Status: closed as duplicate of #35101.

#35101 was opened first and is a superset: same != 0 bypass removal and sourcemap verification, plus seeds section hashes with input_hash and bumps EXPECTED_VERSION to 24.

robobun · 2026-07-22T08:59:43Z

^{Updated 2:05 AM PT - Jul 22nd, 2026}

@robobun, your commit 66c2907 is building: #77675

github-actions · 2026-07-22T09:04:14Z

Found 1 issue this PR may fix:

Security: Fixed-seed non-cryptographic hashes (Wyhash11 / std.hash.Wyhash) are algebraically collidable, enabling trustedDependencies RCE, scoped-registry token leak, and cache poisoning #32741 - This issue identifies RuntimeTranspilerCache as vulnerable to cache poisoning via algebraically collidable fixed-seed wyhash. This PR hardens the transpiler cache integrity checks by removing the zero-hash bypass and adding sourcemap hash verification, directly addressing the cache poisoning vector described in the issue.

If this is helpful, copy the block below into the PR description to auto-close this issue on merge.

Fixes #32741

🤖 Generated with Claude Code

… load Entry::load() skipped the output/esm_record integrity check when the stored hash field was zero, and never verified the sourcemap section at all. An entry's input_hash, input_byte_length and features_hash are all derivable from the public source bytes, so whoever can write a .pile at the expected cache path can prepend arbitrary code to the cached output and zero output_hash to have it accepted unchecked on the next run. Entry::save() has always written all three hashes, so existing on-disk entries already carry the values the loader now compares against.

robobun · 2026-07-22T09:05:22Z

Linked #32741 in the PR body as related but not auto-closing: that issue covers several distinct wyhash-collision vectors (trustedDependencies, scoped-registry, patch cache, MySQL prepared statements, and PoC6's transpiler-cache input_hash collision). This PR only removes the output_hash == 0 bypass in the transpiler cache loader; the per-user keyed hash and @t@ ownership check that would make entries unforgeable are follow-up work.

github-actions · 2026-07-22T09:05:26Z

This PR may be a duplicate of:

transpiler cache: bind section hashes to input and verify unconditionally #35101 - Also removes the zero-hash bypass in Entry::load() to unconditionally verify output/sourcemap/esm_record hashes, fixing the same transpiler cache tampering vulnerability

🤖 Generated with Claude Code

robobun · 2026-07-22T09:05:47Z

Closing in favor of #35101, which was opened first and is a superset of this change: it removes the same != 0 bypass and adds the same sourcemap check, plus seeds each section hash with input_hash (so a forged wyhash(42, tampered_output) no longer matches) and bumps EXPECTED_VERSION accordingly.

coderabbitai

Actionable comments posted: 3

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@test/cli/run/transpiler-cache.test.ts`:
- Around line 364-367: Shorten the comment above the esm_record hash derivation
to three lines or fewer while preserving that it recomputes the hash so the
corrupted record passes integrity validation and reaches deserialization, using
the same wyhash variant and seed as Bun.hash.wyhash.
- Around line 283-291: Shorten the explanatory comment above the transpiler
cache test to three lines or fewer while preserving the essential security
rationale: a zero stored output hash must still trigger verification of cached
output. Keep the issue reference and remove the extended attack-scenario
details.
- Around line 282-330: Update the regression test around the first and second
bunRun calls in rejects a cache entry whose output_hash is zeroed to assert that
both successful runs have exactly empty stderr, matching this file’s bunRun test
convention.

🪄 Autofix (Beta)

Fix all unresolved CodeRabbit comments on this PR:

Push a commit to this branch (recommended)
Create a new PR with the fixes

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: ASSERTIVE

Plan: Pro

Run ID: 2ea6c2a3-11b4-4bdc-a6d4-166eeb5fe73b

📥 Commits

Reviewing files that changed from the base of the PR and between 47597ab and 816e836.

📒 Files selected for processing (2)

src/jsc/RuntimeTranspilerCache.rs
test/cli/run/transpiler-cache.test.ts

coderabbitai · 2026-07-22T09:08:43Z

+test("rejects a cache entry whose output_hash is zeroed", () => {
+  // https://github.com/oven-sh/bun/issues/2829
+  //
+  // The on-disk transpiler cache entry stores a wyhash of the transpiled
+  // output in the header. On load the hash of the output bytes is compared
+  // against that stored value. A stored hash of zero must not skip that
+  // check: anyone who can write to the cache directory can compute the
+  // cache filename for a known source file, prepend arbitrary code to the
+  // cached output, and zero the stored output_hash so the altered output
+  // is accepted without being re-verified.
+
+  // A >=4 KiB source file so it is eligible for the cache.
+  writeFileSync(join(temp_dir, "a.js"), dummyFile(50 * 1024, "output-hash-zero", "clean"));
+
+  // First run transpiles the file and writes the cache entry.
+  const first = bunRun(join(temp_dir, "a.js"), env);
+  expect(first.stdout).toBe("clean");
+  expect(newCacheCount()).toBe(1);
+  const pile = join(cache_dir, readdirSync(cache_dir)[0]);
+
+  // Rewrite the cached output: prepend a payload, then zero output_hash so the
+  // only integrity check on the output bytes is bypassed.
+  const data = readFileSync(pile);
+  const outOff = Number(data.readBigUInt64LE(OUTPUT_BYTE_OFFSET_AT));
+  const outLen = Number(data.readBigUInt64LE(OUTPUT_BYTE_LENGTH_AT));
+  const smOff = Number(data.readBigUInt64LE(SOURCEMAP_BYTE_OFFSET_AT));
+  const smLen = Number(data.readBigUInt64LE(SOURCEMAP_BYTE_LENGTH_AT));
+  const esmOff = Number(data.readBigUInt64LE(ESM_RECORD_BYTE_OFFSET_AT));
+  const esmLen = Number(data.readBigUInt64LE(ESM_RECORD_BYTE_LENGTH_AT));
+
+  const payload = Buffer.from('console.log("PLANTED");\n');
+  const newOut = Buffer.concat([payload, data.subarray(outOff, outOff + outLen)]);
+  const sm = data.subarray(smOff, smOff + smLen);
+  const esm = data.subarray(esmOff, esmOff + esmLen);
+
+  const header = Buffer.from(data.subarray(0, METADATA_SIZE));
+  header.writeBigUInt64LE(BigInt(METADATA_SIZE), OUTPUT_BYTE_OFFSET_AT);
+  header.writeBigUInt64LE(BigInt(newOut.length), OUTPUT_BYTE_LENGTH_AT);
+  header.writeBigUInt64LE(0n, OUTPUT_HASH_AT);
+  header.writeBigUInt64LE(BigInt(METADATA_SIZE + newOut.length), SOURCEMAP_BYTE_OFFSET_AT);
+  header.writeBigUInt64LE(BigInt(METADATA_SIZE + newOut.length + sm.length), ESM_RECORD_BYTE_OFFSET_AT);
+  writeFileSync(pile, Buffer.concat([header, newOut, sm, esm]));
+
+  // Second run must not execute the planted payload. The entry's integrity
+  // check fails, so the cache entry is discarded and the source file is
+  // re-transpiled from scratch.
+  const second = bunRun(join(temp_dir, "a.js"), env);
+  expect(second.stdout).toBe("clean");
+});


📐 Maintainability & Code Quality | 🔵 Trivial | ⚡ Quick win

Correct regression coverage; add empty-stderr assertions per this file's convention.

Logic correctly reproduces the CVE scenario and verifies rejection. However, first.stderr/second.stderr aren't asserted, unlike the established convention in this file.

Based on learnings from this exact file (PR 34721): "prefer asserting exact empty stderr for successful subprocess runs" for bunRun-based tests.

♻️ Proposed addition

const first = bunRun(join(temp_dir, "a.js"), env); expect(first.stdout).toBe("clean"); + expect(first.stderr).toBe(""); expect(newCacheCount()).toBe(1);

const second = bunRun(join(temp_dir, "a.js"), env); expect(second.stdout).toBe("clean"); + expect(second.stderr).toBe(""); });

📝 Committable suggestion

‼️ IMPORTANT
Carefully review the code before committing. Ensure that it accurately replaces the highlighted code, contains no missing lines, and has no issues with indentation. Thoroughly test & benchmark the code to ensure it meets the requirements.

Suggested change

test("rejects a cache entry whose output_hash is zeroed", () => {

// https://github.com/oven-sh/bun/issues/2829

//

// The on-disk transpiler cache entry stores a wyhash of the transpiled

// output in the header. On load the hash of the output bytes is compared

// against that stored value. A stored hash of zero must not skip that

// check: anyone who can write to the cache directory can compute the

// cache filename for a known source file, prepend arbitrary code to the

// cached output, and zero the stored output_hash so the altered output

// is accepted without being re-verified.

// A >=4 KiB source file so it is eligible for the cache.

writeFileSync(join(temp_dir, "a.js"), dummyFile(50 * 1024, "output-hash-zero", "clean"));

// First run transpiles the file and writes the cache entry.

const first = bunRun(join(temp_dir, "a.js"), env);

expect(first.stdout).toBe("clean");

expect(newCacheCount()).toBe(1);

const pile = join(cache_dir, readdirSync(cache_dir)[0]);

// Rewrite the cached output: prepend a payload, then zero output_hash so the

// only integrity check on the output bytes is bypassed.

const data = readFileSync(pile);

const outOff = Number(data.readBigUInt64LE(OUTPUT_BYTE_OFFSET_AT));

const outLen = Number(data.readBigUInt64LE(OUTPUT_BYTE_LENGTH_AT));

const smOff = Number(data.readBigUInt64LE(SOURCEMAP_BYTE_OFFSET_AT));

const smLen = Number(data.readBigUInt64LE(SOURCEMAP_BYTE_LENGTH_AT));

const esmOff = Number(data.readBigUInt64LE(ESM_RECORD_BYTE_OFFSET_AT));

const esmLen = Number(data.readBigUInt64LE(ESM_RECORD_BYTE_LENGTH_AT));

const payload = Buffer.from('console.log("PLANTED");\n');

const newOut = Buffer.concat([payload, data.subarray(outOff, outOff + outLen)]);

const sm = data.subarray(smOff, smOff + smLen);

const esm = data.subarray(esmOff, esmOff + esmLen);

const header = Buffer.from(data.subarray(0, METADATA_SIZE));

header.writeBigUInt64LE(BigInt(METADATA_SIZE), OUTPUT_BYTE_OFFSET_AT);

header.writeBigUInt64LE(BigInt(newOut.length), OUTPUT_BYTE_LENGTH_AT);

header.writeBigUInt64LE(0n, OUTPUT_HASH_AT);

header.writeBigUInt64LE(BigInt(METADATA_SIZE + newOut.length), SOURCEMAP_BYTE_OFFSET_AT);

header.writeBigUInt64LE(BigInt(METADATA_SIZE + newOut.length + sm.length), ESM_RECORD_BYTE_OFFSET_AT);

writeFileSync(pile, Buffer.concat([header, newOut, sm, esm]));

// Second run must not execute the planted payload. The entry's integrity

// check fails, so the cache entry is discarded and the source file is

// re-transpiled from scratch.

const second = bunRun(join(temp_dir, "a.js"), env);

expect(second.stdout).toBe("clean");

});

test("rejects a cache entry whose output_hash is zeroed", () => {

// https://github.com/oven-sh/bun/issues/2829

//

// The on-disk transpiler cache entry stores a wyhash of the transpiled

// output in the header. On load the hash of the output bytes is compared

// against that stored value. A stored hash of zero must not skip that

// check: anyone who can write to the cache directory can compute the

// cache filename for a known source file, prepend arbitrary code to the

// cached output, and zero the stored output_hash so the altered output

// is accepted without being re-verified.

// A >=4 KiB source file so it is eligible for the cache.

writeFileSync(join(temp_dir, "a.js"), dummyFile(50 * 1024, "output-hash-zero", "clean"));

// First run transpiles the file and writes the cache entry.

const first = bunRun(join(temp_dir, "a.js"), env);

expect(first.stdout).toBe("clean");

expect(first.stderr).toBe("");

expect(newCacheCount()).toBe(1);

const pile = join(cache_dir, readdirSync(cache_dir)[0]);

// Rewrite the cached output: prepend a payload, then zero output_hash so the

// only integrity check on the output bytes is bypassed.

const data = readFileSync(pile);

const outOff = Number(data.readBigUInt64LE(OUTPUT_BYTE_OFFSET_AT));

const outLen = Number(data.readBigUInt64LE(OUTPUT_BYTE_LENGTH_AT));

const smOff = Number(data.readBigUInt64LE(SOURCEMAP_BYTE_OFFSET_AT));

const smLen = Number(data.readBigUInt64LE(SOURCEMAP_BYTE_LENGTH_AT));

const esmOff = Number(data.readBigUInt64LE(ESM_RECORD_BYTE_OFFSET_AT));

const esmLen = Number(data.readBigUInt64LE(ESM_RECORD_BYTE_LENGTH_AT));

const payload = Buffer.from('console.log("PLANTED");\n');

const newOut = Buffer.concat([payload, data.subarray(outOff, outOff + outLen)]);

const sm = data.subarray(smOff, smOff + smLen);

const esm = data.subarray(esmOff, esmOff + esmLen);

const header = Buffer.from(data.subarray(0, METADATA_SIZE));

header.writeBigUInt64LE(BigInt(METADATA_SIZE), OUTPUT_BYTE_OFFSET_AT);

header.writeBigUInt64LE(BigInt(newOut.length), OUTPUT_BYTE_LENGTH_AT);

header.writeBigUInt64LE(0n, OUTPUT_HASH_AT);

header.writeBigUInt64LE(BigInt(METADATA_SIZE + newOut.length), SOURCEMAP_BYTE_OFFSET_AT);

header.writeBigUInt64LE(BigInt(METADATA_SIZE + newOut.length + sm.length), ESM_RECORD_BYTE_OFFSET_AT);

writeFileSync(pile, Buffer.concat([header, newOut, sm, esm]));

// Second run must not execute the planted payload. The entry's integrity

// check fails, so the cache entry is discarded and the source file is

// re-transpiled from scratch.

const second = bunRun(join(temp_dir, "a.js"), env);

expect(second.stdout).toBe("clean");

expect(second.stderr).toBe("");

});

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the rest with a brief reason, keep changes minimal, and validate. In `@test/cli/run/transpiler-cache.test.ts` around lines 282 - 330, Update the regression test around the first and second bunRun calls in rejects a cache entry whose output_hash is zeroed to assert that both successful runs have exactly empty stderr, matching this file’s bunRun test convention.

Source: Learnings

robobun assigned Jarred-Sumner Jul 22, 2026

github-actions Bot added the claude label Jul 22, 2026

robobun force-pushed the farm/20db5179/transpiler-cache-reject-zero-output-hash branch from 816e836 to 66c2907 Compare July 22, 2026 09:04

robobun closed this Jul 22, 2026

coderabbitai Bot reviewed Jul 22, 2026

View reviewed changes

Navigation Menu

Uh oh!

transpiler cache: always verify output/sourcemap/esm_record hashes on load - #35102

transpiler cache: always verify output/sourcemap/esm_record hashes on load#35102
robobun wants to merge 1 commit into
mainfrom
farm/20db5179/transpiler-cache-reject-zero-output-hash

robobun commented Jul 22, 2026 •

edited

Loading

Uh oh!

coderabbitai Bot commented Jul 22, 2026 •

edited

Loading

Uh oh!

robobun commented Jul 22, 2026 •

edited

Loading

Uh oh!

robobun commented Jul 22, 2026 •

edited

Loading

Uh oh!

github-actions Bot commented Jul 22, 2026

Uh oh!

robobun commented Jul 22, 2026

Uh oh!

github-actions Bot commented Jul 22, 2026

Uh oh!

robobun commented Jul 22, 2026

Uh oh!

coderabbitai Bot left a comment

Uh oh!

coderabbitai Bot Jul 22, 2026

Uh oh!

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Uh oh!

Conversation

robobun commented Jul 22, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

What

Why it matters

Fix

Scope

Uh oh!

coderabbitai Bot commented Jul 22, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Walkthrough

Changes

Uh oh!

robobun commented Jul 22, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

robobun commented Jul 22, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

github-actions Bot commented Jul 22, 2026

Uh oh!

robobun commented Jul 22, 2026

Uh oh!

github-actions Bot commented Jul 22, 2026

Uh oh!

robobun commented Jul 22, 2026

Uh oh!

coderabbitai Bot left a comment

Choose a reason for hiding this comment

Uh oh!

coderabbitai Bot Jul 22, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

robobun commented Jul 22, 2026 •

edited

Loading

coderabbitai Bot commented Jul 22, 2026 •

edited

Loading

robobun commented Jul 22, 2026 •

edited

Loading

robobun commented Jul 22, 2026 •

edited

Loading