GHSA-8332-j7wj-wjw4

Summary

Three Mongoose code paths copy data into objects by property name without checking whether the property is the object's own or inherited from its prototype:

1. Loading a document from the database (init(), also used by Model.hydrate())
2. Setting a dotted path on a document ($__set, used by new Model(obj), doc.set() and doc.$set())
3. Merging update documents (utils.mergeClone, used by findOneAndReplace(), chained update calls and Query.prototype.merge())

When a key names an inherited property, such as toString or valueOf, or a document internal such as $__schema, Mongoose writes into the shared object behind that property. The data comes from the attacker.

• Denial of service. For example, { "toString": { "call": 1 } } overwrites Object.prototype.toString.call. Mongoose and most other libraries rely on Object.prototype.toString.call(). After the overwrite, every later query in the process throws TypeError: _toString.call is not a function until the process restarts. This includes queries on unrelated models.
• Changing shared schema state. Through a subdocument with strict: false, a path such as profile.$__schema.paths.email.validators reaches the subdocument's schema, which every document shares. Setting it to [] removes the required validator for every document in the process.

Mongoose already blocked __proto__, constructor and prototype, so these bugs cannot add properties to Object.prototype itself. They can still overwrite methods and properties of shared built-in functions and of Mongoose's internal objects.

Details

1. Loading documents: init()

if (!schemaType && utils.isPOJO(value)) {
  // assume nested object
  if (!doc[i]) { doc[i] = {}; ... }
  init(self, value, doc[i], opts, path + '.');
}

For a stored field named toString, doc[i] is the inherited Object.prototype.toString. It is truthy, so init() recurses into that function and copies the stored subfields onto it.

This runs for every document Mongoose loads: find(), findOne(), cursors, populated documents and Model.hydrate(). It runs whatever the strict setting, and it applies to top-level and nested paths alike. lean() queries are not affected.

A malicious document can get into the collection in several ways:

• through a model with strict: false, for example new Model(req.body).save();
• through another application or the raw MongoDB driver;
• through any update that writes arbitrary keys.

2. Setting dotted paths: $__set

When it walks a dotted path, $__set read obj[parts[i]] and descended into whatever it found. That included inherited functions, such as Object.prototype.toString on a plain object or Document.prototype.toObject on a subdocument. On subdocuments it also included non-function internals that are prototype getters, such as $__schema, schema, $__middleware and $__schemaTypeOptions. Those point at objects that every document of the model shares.

Attacker-controlled input reaches this path through:

• A Mixed path on a default, strict schema: new Model({ "mixed.toString.call": 1 }), or doc.set({ "mixed.toString.call": 1 }). Nothing needs to be saved.
• Any root path on a schema with strict: false: for example doc.set('toString.call', 1).
• A subdocument with strict: false, when the application passes a user-controlled path to doc.set(path, value): for example doc.set('profile.$__schema.paths.email.validators', []).

3. Merging updates: utils.mergeClone

if (typeof to[key] === 'undefined') {
  to[key] = clone(fromObj[key]);
} else if (isObject(val)) {
  mergeClone(to[key], val);   // to[key] may be an inherited built-in
}

to[key] was read without an own-property check, so a key such as toString recursed into Object.prototype.toString. This affects:

• Model.findOneAndReplace(filter, replacement). It always merges the replacement into the query's update, so a single call with an untrusted replacement is enough.
• Update calls chained onto a query that already has an update. For example Model.updateOne(filter, { $set: {...} }).updateOne({}, untrustedUpdate), the same pattern with findOneAndUpdate(), or middleware that calls this.updateOne(...).
• Query.prototype.merge() with a source query whose update came from untrusted input.

Proof of concept

const mongoose = require('mongoose');
const { Schema } = mongoose;

// 2. Setting dotted paths: a Mixed path on a default strict schema, nothing saved
const Item = mongoose.model('Item', new Schema({ name: String, meta: {} }));
new Item(JSON.parse('{"meta.toString.call": 1}'));  // e.g. new Item(req.body)
console.log(typeof Object.prototype.toString.call); // 'number'
// From here on, every query in the process fails:
// TypeError: _toString.call is not a function

// 1. Loading documents: a stored field named after an Object.prototype method
const Loose = mongoose.model('Loose', new Schema({ name: String }, { strict: false }));
await new Loose(JSON.parse('{"name":"evil","toString":{"call":1}}')).save();
await Loose.findOne({ name: 'evil' });              // any later read of this document
await OtherModel.findOne();                         // TypeError: _toString.call is not a function

// 3. Merging updates: a single findOneAndReplace() call with an untrusted replacement
const Doc = mongoose.model('Doc', new Schema({ name: String }));
Doc.findOneAndReplace({}, JSON.parse('{"toString":{"call":1}}'));
console.log(typeof Object.prototype.toString.call); // 'number'

// 2. Setting dotted paths: changing the shared subdocument schema
const User = mongoose.model('User', new Schema({
  profile: new Schema({ email: { type: String, required: true } }, { strict: false })
}));
// e.g. JSON-patch style endpoint: doc.set(req.body.path, req.body.value)
new User({ profile: { email: 'a@b.c' } }).set('profile.$__schema.paths.email.validators', []);
await new User({ profile: {} }).validate();         // resolves: `required` removed for every document

Impact

• Denial of service (paths 1, 2 and 3). One request overwrites Object.prototype.toString.call. After that, every query in the process fails, and so does much other code, until the process restarts. Ways in:

  • new Model(req.body) on any schema with a Mixed path;
  • findOneAndReplace() with an untrusted replacement;
  • reading one stored document that contains a field such as { "toString": { "call": 1 } }.

  In the stored-document case, the attacker does not need any write path through Mongoose at the moment of the attack. Any later read of that document triggers the overwrite. If the document is read on startup or by a scheduled job, the application can be kept down.

• Changes to shared state (path 2, strict: false subdocuments). An attacker who controls the path passed to doc.set(path, value) can change a subdocument schema that every document shares. For example, they can remove validators or change schema options for every later document in the process.

• Properties on built-in functions. More generally, an attacker can attach properties they choose to built-in functions such as Object.prototype.toString, Object.prototype.valueOf and Function.prototype.call. Any other code that reads properties of those functions sees the attacker's values.