GHSA-r292-9mhp-454m

ADVISORY - github

Summary

Summary

node-tar (npm tar) contains an uncontrolled-recursion stack-exhaustion DoS in the internal mapHas helper used by filesFilter. When a consumer calls tar.t(...) or tar.x(...) with a non-empty member-selection list, node-tar installs a filter that closes over the recursive mapHas (src/list.ts:33-44). mapHas walks an entry path upward one path.dirname() call per recursion with no segment cap. A single crafted tar with a GNU-L (or PAX-x) long-path header can deliver a path of tens of thousands of /-separated segments (up to maxMetaEntrySize = 1 MiB). The recursion overflows the call stack, throwing an uncatchable RangeError that terminates the Node process on async/streaming consumers.

Root Cause

filesFilter (src/list.ts:27-51) is installed whenever a caller passes a member-selection list (src/list.ts:119-122, src/extract.ts:55-57). Its filter is invoked at src/parse.ts:253 (entry.ignore = entry.ignore || !this.filter(entry.path, entry)) inside Parser[CONSUMEHEADER] — and crucially outside the only try/catch in that method (which wraps new Header at src/parse.ts:179-183). mapHas recurses once per path segment with no depth limit. The Unpack maxDepth guard (src/unpack.ts:342, in [CHECKPATH]) only runs on the 'entry' event, which fires after CONSUMEHEADER has already invoked the filter — so the stack overflows before any depth guard executes. tar.t (list) has no maxDepth at all.

Impact

Unauthenticated, remotely-triggerable denial of service: a ~188-byte gzip (≈26 KB tar) crashes any service that lists or extracts selected members from an untrusted archive (package registries, CI artifact/cache restore, upload processors). On async (await tar.t(...)/tar.x(...)) and streaming/pipe consumers the RangeError escapes the promise as an uncaughtException and terminates the process — standard defensive try/catch around the async call does NOT prevent it. (The synchronous API is catchable; the async/stream paths — the dominant server pattern — are not.)

Proof of Concept

// Build a tar whose single entry has a GNU-L long path of ~12,000 "a/" segments (~26 KB),
// gzip it (≈188 bytes), then have a consumer list/extract with member selection:
const tar = require('tar');
await tar.t({ file: 'evil.tar.gz', gzip: true }, ['some-member']); // -> RangeError, process exit

Empirically reproduced on Node v24.18.0 against built dist/commonjs of node-tar 7.5.20: 188-byte gzip → 26,112-byte tar (12,000 segments) → uncaught RangeError: Maximum call stack size exceeded → process exit. A control run with no member-selection list (filter not installed) parses cleanly (exit 0), isolating mapHas as the sole cause.

Attack Chain

  1. Entry. Attacker crafts a tar with a GNU L (or PAX x) long-path header whose body is "a/"×12000 (26 KB), followed by a normal file entry.
    • Guard: maxMetaEntrySize caps the meta body at 1 MiB (src/parse.ts:241).
    • Bypass proof: 26 KB ≪ 1 MiB → accepted (verified: 26 KB archive parsed up to the filter).
  2. Trigger. Victim service calls tar.t({file},[sel]) or tar.x({file,cwd},[sel]) (member selection — a documented, common API).
    • Guard: Unpack.maxDepth (default 1024) at src/unpack.ts:342; decompression-ratio guard.
    • Bypass proof: maxDepth lives in [CHECKPATH] on the 'entry' event, which fires after CONSUMEHEADER's filter call — the crash occurs before it (extract exits 1 with default maxDepth). tar.t has no maxDepth. Ratio is ~139× (trivial); no total-bytes cap applies to the uncompressed meta body.
  3. Sink. this.filter(entry.path)mapHas recurses once per / segment (src/list.ts:39).
    • Guard: try/catch in CONSUMEHEADER.
    • Bypass proof: the only try/catch wraps new Header (src/parse.ts:179-183); the this.filter(...) call at src/parse.ts:253 is outside it. The RangeError propagates out of the stream write/'data' path → uncaught exception (verified: process.on('uncaughtException') fires; async await+try/catch does NOT intercept).
  4. Impact. Node process termination; a 188-byte gzip crashes any consumer that lists/extracts selected members from untrusted archives.

Bypass Evidence

  • mapHas recursion is member-name-independent: the crash fires even when the requested members do not match the malicious entry path — the attacker only needs the consumer to use member selection.
  • Standalone mapHas overflows at 20k–30k segments; on the real streaming path (atop write → CONSUMECHUNK → CONSUMEHEADER → filter) it crashes at ≤8k segments (finder's ~12k estimate is accurate for the reachable path).
  • Control (no member list → no filter) parses cleanly (exit 0), isolating mapHas.

Affected Versions

<= 7.5.20 (npm tar). mapHas present verbatim on tag v7.5.20 (latest GitHub release and npm dist-tag latest); no segment/depth cap in src/list.ts or the CONSUMEHEADER filter path; HEAD == 7.5.20, no unreleased fix.

Suggested Fix

Rewrite mapHas iteratively (walk dirname in a while loop with a segment/visited cap), or enforce a hard path-segment limit in Header/Parser independent of maxMetaEntrySize, applied before any per-entry filter runs.

Dedup Note

Distinct from CVE-2024-28863 / GHSA-f5x3-32g6-qm9j "lack of folders depth validation" (that bounds mkdir recursion during extraction via maxDepth in Unpack[CHECKPATH] on the 'entry' event — a different sink, code path, and fix; runs after the filter and does not apply to tar.t). Also distinct from the PAX NUL/numeric-path crash advisories (improper-input-to-fs / type confusion, not recursion) and the gzip-bomb advisory (resource exhaustion on disk writes). None touch list.ts/filesFilter/mapHas or require member selection.


Reported by zx (Jace) — GitHub: @manus-use

Common Weakness Enumeration (CWE)

ADVISORY - github

Uncontrolled Resource Consumption

Uncontrolled Recursion


GitHub

CREATED

UPDATED

EXPLOITABILITY SCORE

3.9

EXPLOITS FOUND
-
COMMON WEAKNESS ENUMERATION (CWE)

CVSS SCORE

5.3medium
PackageTypeOS NameOS VersionAffected RangesFix Versions
tarnpm--<=7.5.207.5.21

CVSS:3 Severity and metrics

The CVSS metrics represent different qualitative aspects of a vulnerability that impact the overall score, as defined by the CVSS Specification.

The vulnerable component is bound to the network stack, but the attack is limited at the protocol level to a logically adjacent topology. This can mean an attack must be launched from the same shared physical (e.g., Bluetooth or IEEE 802.11) or logical (e.g., local IP subnet) network, or from within a secure or otherwise limited administrative domain (e.g., MPLS, secure VPN to an administrative network zone). One example of an Adjacent attack would be an ARP (IPv4) or neighbor discovery (IPv6) flood leading to a denial of service on the local LAN segment (e.g., CVE-2013-6014).

Specialized access conditions or extenuating circumstances do not exist. An attacker can expect repeatable success when attacking the vulnerable component.

The attacker is unauthorized prior to attack, and therefore does not require any access to settings or files of the vulnerable system to carry out an attack.

The vulnerable system can be exploited without interaction from any user.

An exploited vulnerability can only affect resources managed by the same security authority. In this case, the vulnerable component and the impacted component are either the same, or both are managed by the same security authority.

There is no loss of confidentiality.

There is no loss of trust or accuracy within the impacted component.

Performance is reduced or there are interruptions in resource availability. Even if repeated exploitation of the vulnerability is possible, the attacker does not have the ability to completely deny service to legitimate users. The resources in the impacted component are either partially available all of the time, or fully available only some of the time, but overall there is no direct, serious consequence to the impacted component.