On Node.js 24 and newer, vm2 can expose the host node:test module to sandboxed NodeVM code when the embedder explicitly allows the node:test builtin. Sandbox code can reach that module through require('node:node:test') and call run() with attacker-controlled execArgv.
node:test.run() starts a separate Node process for process-isolated test execution and forwards the supplied execArgv values to that process. Supplying --eval=<JavaScript> therefore executes arbitrary JavaScript in an unrestricted host Node process, outside the NodeVM sandbox.
The PoC confirms that direct sandbox imports of fs, child_process, module, and process remain denied before the spawned process imports host fs and writes a harmless marker.
vm2>=3.9.6, <=3.11.53.11.5v24.18.0module.builtinModules does not expose the scheme-only node:test entry thererequire: {
builtin: ['node:test'],
external: false
}
The lower version boundary was tested directly: vm2@3.9.5 blocks require('node:node:test'), while vm2@3.9.6 permits the exploit path. Representative releases through 3.11.5 were also reproduced.
The issue is a combination of builtin admission, generic host passthrough, and prefix normalization:
module.builtinModules includes the scheme-only key node:test.lib/builtin.js builds BUILTIN_MODULES from that array. The family-based DANGEROUS_BUILTINS protection does not include test, so node:test remains eligible.node:test, addDefaultBuiltin() stores it through the generic loader:builtins.set(key, special ? special : vm => vm.readonly(hostRequire(key)));
lib/setup-node-sandbox.js, requireImpl() strips one node: prefix before builtin lookup:if (localStringPrototypeStartsWith(filename, 'node:')) {
id = localStringPrototypeSlice(filename, 5);
nmod = loadBuiltinModule(id);
}
node:node:test is normalized to the stored key node:test and receives a readonly proxy to the host module.node:test.run() safe. Calls are forwarded to the host implementation, which accepts attacker-controlled execArgv for a newly spawned Node process.--eval=<attacker JavaScript> runs outside vm2 and has normal host builtin access.The doubled prefix is the reachability mechanism, but the security boundary failure is broader: the generic host-passthrough loader treats the test builtin family as safe even though its run() API can launch unrestricted Node processes.
From the poc directory:
npm ci --ignore-scripts --no-audit --no-fund
node repro.js
Expected successful result on Node.js 24+ includes:
{
"vm2Version": "3.11.5",
"nodeVersion": "v24.18.0",
"markerExists": true,
"childIsDistinctProcess": true,
"marker": {
"hostCodeExecution": true
}
}
The PoC writes only host-rce-marker.json in its own directory and does not invoke a shell, contact a network service, or access third-party data.
An attacker who is intentionally permitted to execute untrusted JavaScript in the affected NodeVM configuration can escape the sandbox and execute arbitrary JavaScript under the embedder's operating-system identity.
This provides the spawned process with the host user's filesystem, environment, network, and process-execution permissions. It can therefore result in complete confidentiality, integrity, and availability impact for the hosting service.
Treat the normalized test builtin family as dangerous before wildcard expansion and explicit builtin registration.
For example, add test to DANGEROUS_BUILTINS so the existing prefix and family checks reject both node:test and node:test/reporters:
const DANGEROUS_BUILTINS = new Set([
// existing entries
'test'
]);
If test helpers must be exposed, provide a sandbox-local wrapper through mock or override that does not expose run(), process isolation, execArgv, or other host process controls.
Recommended regression cases:
builtin: ['node:test']require('node:test')require('node:node:test')node:test/reporters and prefixed variants--eval, --require, or --import through test-runner process options
vm2-node-test-ghsa-submission.zip{
"cwe_ids": [
"CWE-693"
],
"github_reviewed": true,
"github_reviewed_at": "2026-10-01T15:41:31Z",
"nvd_published_at": null,
"severity": "CRITICAL"
}