GHSA-67c8-pqhq-4rmx: Npm/piscina vulnerability

Summary

A prototype-pollution gadget in ThreadPool.options allows an attacker who can pollute Object.prototype to execute arbitrary code in Piscina worker threads, invoke arbitrary functions during task scheduling, or inject environment variables into workers. The root cause is that ThreadPool.options is created as a plain object inheriting from Object.prototype, so any option without an explicit default in kDefaultOptions can be supplied via the prototype chain.

Details

In src/index.ts the ThreadPool constructor builds the resolved options object as a plain object:

ts this.options = { ...kDefaultOptions, ...options, filename, maxQueue: 0 }

Because this object has Object.prototype as its prototype, reads for properties that are not own properties of the object and are not present in kDefaultOptions fall back to Object.prototype. This means a prototype-pollution primitive (e.g. from a vulnerable merge() or JSON.parse merge elsewhere in the application) can inject values for execArgv, env, loadBalancer, argv, workerData, resourceLimits, niceIncrement, closeTimeout, recordTiming, stricterFIFO, workerHistogram, and trackUnmanagedFds.

The most serious gadget is execArgv, which is passed directly to new Worker(..., { execArgv }). An attacker can set Object.prototype.execArgv = ['--require', '/tmp/attacker.js'], causing every worker to preload and execute the attacker-controlled module on startup.

This issue survived the fix for GHSA-x9g3-xrwr-cwfg / CVE-2026-55388 ("Prototype Pollution Gadget → RCE via inherited options.filename"). That advisory hardened the Piscina constructor's filename read and run()'s filename/name reads, but ThreadPool.options itself was not created with a null prototype. The same class of attack is therefore still possible against any option without an explicit default in kDefaultOptions.

PoC

js import { resolve } from 'node:path' import Piscina from 'piscina'

Object.prototype.execArgv = ['--require', '/tmp/attacker.js']

const pool = new Piscina({ filename: resolve(import.meta.dirname, 'worker.js'), minThreads: 1, maxThreads: 1, })

await pool.run(1)

/tmp/attacker.js is executed in the worker on startup. A full reproduction repository with execArgv, loadBalancer, and env vectors is available at https://github.com/Fcmam5/piscina-pp-poc.

Impact

- Remote Code Execution: via execArgv (arbitrary --require module preloaded in every worker on spawn). - Arbitrary code execution in the main thread: via loadBalancer, an attacker-supplied function that is called during task scheduling. - Environment/CLI option injection: via env, which is passed to each worker constructor. - Denial of Service / unexpected behavior: via other reachable options such as workerData, resourceLimits, niceIncrement, closeTimeout, recordTiming, etc.

Anyone using Piscina in an application where Object.prototype can be polluted (e.g. through a dependency with a prototype-pollution vulnerability) is impacted.