CVE-2026-92944

Reporter

• Name or handle: [YMsora]
• Report date: 2026-08-14

The latest published vm2 release, 3.11.5, and the current main branch are vulnerable to a sandbox escape when used on Node.js 26. An ordinary fulfilled Promise created by an async function can retain an attacker-controlled constructor[Symbol.species] across Promise.prototype.finally().

vm2 installs wrappers on the intrinsic Promise.prototype.then and catch methods. On Node.js 26 / V8 14.6, V8's SetPrototypeProperties path updates these existing data properties without invalidating the PromiseThenLookupChain protector. Promise.prototype.finally() subsequently trusts the stale protector and uses an internal InvokeThen fast path that calls the original native then, bypassing vm2's wrapper and its resetPromiseSpecies(this) hardening.

The attacker-controlled species constructor therefore supplies the resolve and reject functions used by a native Promise reaction. A calibrated stack overflow at that native reaction boundary yields a raw host-realm RangeError to the attacker-controlled reject function. Its constructor chain reaches the host Function constructor and consequently the host process object.

The attached proof is non-destructive: it reads only process.version. It does not execute commands, access files, or perform network requests.

Latest-version status

Verified on 2026-08-14:

• npm latest: [email protected]
• [email protected] publication time: 2026-05-18T15:37:26.138Z
• npm gitHead: 7a1f5100b96f48d34e0fe104ab37c0acc5944f92
• npm tarball: https://registry.npmjs.org/vm2/-/vm2-3.11.5.tgz
• npm integrity: sha512-RSrkBiwrj6FRU+QdqNs6KG0XdlvJCjpQ4GXiqmMbrhmwfu5k/XIMpAer0L8f6iuf0uJ3a4T1xJN126Q8yf0VIA==
• GitHub default branch main HEAD: the same commit, also tagged v3.11.5
• latest formal Node.js release: v26.7.0, V8 14.6.202.34

The npm tarball and current GitHub main contain the same relevant lib/setup-sandbox.js Git blob. There is no unpublished fix on main at the time of this report.

Permanent source reference:

https://github.com/patriksimek/vm2/blob/7a1f5100b96f48d34e0fe104ab37c0acc5944f92/lib/setup-sandbox.js#L345-L390

Affected configurations

vm2 versions tested on Node.js 26.7.0

This report therefore confirms vm2 3.10.2 through 3.11.5 as affected when used with the vulnerable runtime. Earlier vm2 versions are not claimed: they may be independently vulnerable through older published issues, which would not establish this root cause.

Node.js versions tested with stock vm2 3.11.5

The vulnerable behavior was reproduced on Linux/musl, Linux/glibc, and Windows x64. It is not Alpine-specific.

Configuration requirements

The default configuration is affected:

const vm = new VM();

Default new VM() was verified 3/3 on Node.js 26.7.0 and vm2 3.11.5 with the operating system's default Node stack size.

The escape also remains reproducible with stronger settings:

const vm = new VM({
  allowAsync: true,
  eval: false,
  wasm: false,
  timeout: 5000
});

Therefore the exploit does not require:

• WebAssembly or JSPI;
• dynamic evaluation being enabled;
• Buffer;
• an exposed host object;
• a custom Promise supplied by the embedder;
• a specific operating system; or
• any challenge-specific code.

The demonstrated producer uses an async function, so async support must be available. This is vm2's default.

The V8 PromiseThenLookupChain protector must still be intact when vm2 installs its wrappers. A fresh/default Node.js process satisfies this condition. An unrelated earlier mutation that correctly invalidates the protector can make this exact path safe, but that is not a documented vm2 mitigation and is not present in the default setup.

Reproduction

The attachment poc.js uses only the published npm package and prints a benign host-version marker.

mkdir vm2-node26-repro
cd vm2-node26-repro
npm init -y
npm install --ignore-scripts --no-audit --no-fund [email protected]
cp /path/to/poc.js .
node poc.js

Alternatively, with the official Node.js 26.7.0 container:

docker run --rm -v "$PWD:/poc:ro" -w /tmp node:26.7.0-bookworm-slim sh -lc '
  npm init -y >/dev/null 2>&1 &&
  npm install --ignore-scripts --no-audit --no-fund [email protected] >/dev/null 2>&1 &&
  cp /poc/poc.js . &&
  node poc.js
'

Representative output:

{
  "node": "v26.7.0",
  "v8": "14.6.202.34-node.28",
  "vm2": "3.11.5",
  "config": "default",
  "result": "HOST",
  "hostVersion": "v26.7.0",
  "speciesCalls": 1,
  "localError": false,
  "localRangeError": false
}

The calibrated depth varies with the platform and process layout; this is expected. The PoC searches the boundary instead of relying on a fixed depth.

To verify that disabling eval and WebAssembly is not sufficient:

STRICT=1 node poc.js

Technical root cause

1. vm2 replaces the intrinsic Promise.prototype.then and catch methods with wrappers that sanitize callbacks and execute resetPromiseSpecies(this).
2. These two prototype writes are consecutive direct assignments in lib/setup-sandbox.js.
3. V8 14.6 enables the proto_assign_seq_opt optimization, combining this assignment sequence into SetPrototypeProperties.
4. In the Node.js 26 implementation, the existing-data-property branch calls Object::SetDataProperty(&it, value) without first calling it.UpdateProtector().
5. The JavaScript property now contains vm2's wrapper, but the V8 PromiseThenLookupChain protector incorrectly remains valid.
6. Promise.prototype.finally() performs an internal InvokeThen. Because the protector appears valid, V8 directly selects the native Promise then instead of performing the observable property lookup that would reach vm2's wrapper.
7. The wrapper never executes, so the attacker's own constructor[Symbol.species] is not reset before the native then creates its result capability.
8. A native Promise reaction calls the attacker-provided capability resolve function. At a calibrated stack boundary, V8 creates a host-realm RangeError and passes it to the attacker-provided capability reject function without vm2 conversion.
9. error.constructor.constructor is consequently the host Function, even when the VM was configured with eval:false.

The root-cause control is strong:

node --no-proto-assign-seq-opt poc.js

On Node.js 26.7.0 this changes the result from HOST to SAFE (3/3). A separate semantic probe also changes from wrapperCalls=0 to wrapperCalls=1.

Relationship to previous advisories

This report intentionally discloses the overlap with prior work:

CVE-2026-22709 / GHSA-99p7-6v5w-7xg8

That issue concerned bypassing callback sanitization on intrinsic Promises returned by async functions. vm2's fix added/strengthened the globalPromise.prototype.then and catch wrappers. The current issue is different: on V8 14.6, finally() trusts a stale protector and bypasses those installed wrappers at the engine fast path.

CVE-2026-47208 / GHSA-76w7-j9cq-rx2j

That issue used a missing resetPromiseSpecies call in localPromise's rejection-swallowing tail. It is marked fixed in 3.11.4. The current entry point is the intrinsic async Promise plus finally()/InvokeThen; the PoC succeeds against 3.11.4 and 3.11.5.

CVE-2026-47210 / GHSA-6j2x-vhqr-qr7q

That issue required a JSPI-backed Promise and WebAssembly APIs. Its 3.11.4 fix removes the relevant JSPI surface. The current PoC uses an ordinary fulfilled async-function Promise, works with wasm:false, and succeeds against 3.11.5.

Public V8 fix

The underlying V8 protector bookkeeping bug is already public and fixed upstream:

• data-property branch: https://github.com/v8/v8/commit/7b26e81087ec88c287d9a4812f969d412f32774e
• follow-up lazy-accessor branch: https://github.com/v8/v8/commit/4635ddd85839461c8ad791990023a477efefe7bd

This report does not claim discovery of a new V8 bug. It reports a previously undocumented, still-unpatched vm2 sandbox escape created by the interaction between that V8 bug and vm2's Promise hardening. The escape affects vm2's latest release and current main branch.

Suggested remediation

1. Install the intrinsic then and catch wrappers through an operation that reliably invalidates V8's Promise lookup-chain protector, such as an appropriate Reflect.defineProperty/Object.defineProperty path, instead of the optimizable consecutive direct-assignment sequence.
2. Wrap the intrinsic Promise.prototype.finally entry point and execute resetPromiseSpecies(this) before delegating to the cached native finally implementation.
3. Add a build/runtime regression test using an intrinsic Promise from (async () => 1)(): an attacker-controlled own species must not be constructed by p.finally().
4. Until a vm2 release is available, document Node.js 26 as affected or fail closed on the vulnerable runtime range.
5. Coordinate with Node.js to backport the existing V8 protector fixes to the Node.js 26 release line.

Verified controls show that wrapping finally before calling the cached native implementation blocks the tested direct variants, while the upstream V8 fix also restores correct wrapper invocation.

Credit is requested as: YMsora. https://github.com/YMs0ra

Attachments

vm2-node26-finally-private-report.zip