Running Untrusted JavaScript Code

IMPORTANT: This is about running JavaScript and TypeScript code only. That being said, the writing might also be the direction to run other code in other languages.

Allowing users to execute their code within your application opens up a world of customization and functionality, yet it also exposes your platform to significant security threats.

Given that it is user code, everything is expected, from halting the servers (it could be infinity loops) to stealing sensitive information.

This article will explore various strategies to mitigate run user code, including Web Workers, static code analysis, and more…

You should care

There are many scenarios where you need to run user-provided code, ranging from collaborative development environments like CodeSandbox and StackBiltz to customizable API platforms like January. Even code playgrounds are susceptible to risks.

Namely, the two essential advantages of safely running user-provided code are:

1. Gaining your user’s trust: Even if the user is trustworthy, they may execute code copied from other intently bad people.
2. Secure your environment: the last thing you need is a piece of code halting your server. Think while (true) {}

Define “Sensitive information”

Running user code isn’t harmful until you’re concerned that this might subject some data to be stolen. Whatever data you’re concerned about will be considered sensitive information. For instance, in most cases, JWT is sensitive information (perhaps when used as an authentication mechanism)

What could go wrong

Consider the potential risks of JWT stored in cookies sent with every request. A user could inadvertently trigger a request that sends the JWT to a malicious server, and...

• Cross-Site Scripting (XSS).
• Denial of Service (DoS) attacks.
• Data exfiltration. Without proper safeguards, these threats can compromise the integrity and performance of your application.

Methods

The Evil Eval

The simplest of all, yet the riskiest.

eval('console.log("I am dangerous!")');

When you run this code, it logs that message. Essentially, eval is a JS interpreter capable of accessing the global/window scope.

const res = await eval('fetch(`https://jsonplaceholder.typicode.com/users`)');
const users = await res.json();

This code uses fetch which is defined in the global scope. The interpreter doesn’t know about it, but since eval can access a window, it knows. That implies that running an eval in the browser is different from running it in a server environment or worker.

eval(`document.body`);

How about this...

eval(`while (true) {}`);

This code will halt the browser tab. You might ask why a user would do this to themselves. Well, they might be copying code from the Internet. That's why it's preferred to do static analysis with/or time-box the execution.

You might want to check MDN Docs about eval

Time box execution can be done by running the code in a web worker and using setTimeout to limit the execution time.

async function timebox(code, timeout = 5000) {
  const worker = new Worker('user-runner-worker.js');
  worker.postMessage(code);

  const timerId = setTimeout(() => {
    worker.terminate();
    reject(new Error('Code execution timed out'));
  }, timeout);

  return new Promise((resolve, reject) => {
    worker.onmessage = event => {
      clearTimeout(timerId);
      resolve(event.data);
    };
    worker.onerror = error => {
      clearTimeout(timerId);
      reject(error);
    };
  });
}

await timebox('while (true) {}');

Function Constructor

This is similar to eval but it’s a bit safer since it can’t access the enclosing scope.

const userFunction = new Function('param', 'console.log(param);');
userFunction(2);

This code will log 2.

Note: The second argument is the function body.

The function constructor can’t access the enclosing scope so that the following code will throw an error.

function fnConstructorCannotUseMyScope() {
  let localVar = 'local value';
  const userFunction = new Function('return localVar');
  return userFunction();
}

But it can access the global scope so the fetch example from above works.

WebWorker

You can run “Function Constructor and eval on a WebWorker, which is a bit safer due to the fact that there is no DOM access.

To put more restrictions in place, consider disallowing using global objects like fetch, XMLHttpRequest, sendBeacon Check this writing about how you can do that.

Isolated-VM

Isolated-VM is a library that allows you to run code in a separate VM (v8's Isolate interface)

import ivm from 'isolated-vm';

const code = `count += 5;`;

const isolate = new ivm.Isolate({ memoryLimit: 32 /* MB */ });
const script = isolate.compileScriptSync(code);
const context = isolate.createContextSync();

const jail = context.global;
jail.setSync('log', console.log);

context.evalSync('log("hello world")');

This code will log hello world

WebAssembly

This is an exciting option as it provides a sandboxed environment to run code. One caveat is that you need an environment with Javascript bindings. However, an interesting project called Extism facilitates that. You might want to follow their tutorial.

What is fascinating about it is that you’ll use eval to run the code, but given WebAssembly’s nature, DOM, network, file system, and access to the host environment are not possible (although they might differ based on the wasm runtime).

function evaluate() {
  const { code, input } = JSON.parse(Host.inputString());
  const func = eval(code);
  const result = func(input).toString();
  Host.outputString(result);
}

module.exports = { evaluate };

You'll have to compile the above code first using Extism, which will output a Wasm file that can be run in an environment that has Wasm-runtime (browser or node.js).

const message = {
  input: '1,2,3,4,5',
  code: `
        const sum = (str) => str
          .split(',')
          .reduce((acc, curr) => acc + parseInt(curr), 0);
        module.exports = sum;
`,
};

// continue running the wasm file

Docker

We're now moving to the server-side, Docker is a great option to run code in an isolation from the host machine. (Beware of container escape)

You can use dockerode to run the code in a container.

import Docker from 'dockerode';
const docker = new Docker();

const code = `console.log("hello world")`;
const container = await docker.createContainer({
  Image: 'node:lts',
  Cmd: ['node', '-e', code],
  User: 'node',
  WorkingDir: '/app',
  AttachStdout: true,
  AttachStderr: true,
  OpenStdin: false,
  AttachStdin: false,
  Tty: true,
  NetworkDisabled: true,
  HostConfig: {
    AutoRemove: true,
    ReadonlyPaths: ['/'],
    ReadonlyRootfs: true,
    CapDrop: ['ALL'],
    Memory: 8 * 1024 * 1024,
    SecurityOpt: ['no-new-privileges'],
  },
});

Keep in mind that you need to make sure the server has docker installed and running. I'd recommend having a separate server dedicated only to this that acts as a pure-function server.

Moreover, you might benefit from taking a look at sysbox, a VM-like container runtime that provides a more secure environment. Sysbox is worth it, especially if the main app is running in a container, which means that you'll be running Docker in Docker.

This was the method of choice at January but soon enough, the language capabilities mandated more than passing the code through the container shell. Besides, for some reason, the server memory spikes frequently; we run the code inside self-removable containers on every 1s debounced keystroke. (You can do better!)

Other options

• Web Containers
• MicroVM (Firecraker)
• Deno subhosting
• Wasmer
• ShadowRealms

Safest option

I'm particularly fond of Firecracker, but it’s a bit of work to set up, so if you cannot afford the time yet, you want to be on the safe side, do a combination of static analysis and time-boxing execution. You can use esprima to parse the code and check for any malicious act.

How to run TypeScript code?

Well, same story with one (could be optional) extra step: Transpile the code to JavaScript before running it. Simply put, you can use esbuild or typescript compiler, then continue with the above methods.

async function build(userCode: string) {
  const result = await esbuild.build({
    stdin: {
      contents: `${userCode}`,
      loader: 'ts',
      resolveDir: __dirname,
    },
    inject: [
      // In case you want to inject some code
    ],
    platform: 'node',
    write: false,
    treeShaking: false,
    sourcemap: false,
    minify: false,
    drop: ['debugger', 'console'],
    keepNames: true,
    format: 'cjs',
    bundle: true,
    target: 'es2022',
    plugins: [
      nodeExternalsPlugin(), // make all the non-native modules external
    ],
  });
  return result.outputFiles![0].text;
}

Notes:

• Rust-based bundlers usually offer a web assembly version, which means you can transpile the code in the browser. Esbuild does have a web assembly version.
• Don't include user specified imports into the bundle unless you've allow-listed them.

Additionally, you can avoid transpiling altogether by running the code using Deno or Bun in a docker container since they support TypeScript out of the box.

Conclusion

Running user code is a double-edged sword. It can provide a lot of functionality and customization to your platform, but it also exposes you to significant security risks. It’s essential to understand the risks and take appropriate measures to mitigate them and remember that the more isolated the environment, the safer it is.

References

• January instant compilation
• Running untrusted JavaScript in Node.js
• How do languages support executing untrusted user code at runtime?
• Safely Evaluating JavaScript with Context Data