Google Chrome Extensions

Using eval in Chrome Extensions. Safely.

Chrome's extension system enforces a fairly strict default Content Security Policy (CSP) . The policy restrictions are straightforward: script must be moved out-of-line into separate JavaScript files, inline event handlers must be converted to use addEventListener, and eval() is disabled. Chrome Apps have an even more strict policy, and we're quite happy with the security properties these policies provide.

We recognize, however, that a variety of libraries use eval() and eval-like constructs such as new Function() for performance optimization and ease of expression. Templating libraries are especially prone to this style of implementation. While some (like Angular.js) support CSP out of the box, many popular frameworks haven't yet updated to a mechanism that is compatible with extensions' eval-less world. Removing support for that functionality has therefore proven more problematic than expected for developers.

This document introduces sandboxing as a safe mechanism to include these libraries in your projects without compromising on security. For brevity, we'll be using the term extensions throughout, but the concept applies equally to applications.

Why sandbox?

eval is dangerous inside an extension because the code it executes has access to everything in the extension's high-permission environment. A slew of powerful chrome.* APIs are available that could severely impact a user's security and privacy; simple data exfiltration is the least of our worries. The solution on offer is a sandbox in which eval can execute code without access either to the extension's data or the extension's high-value APIs. No data, no APIs, no problem.

We accomplish this by listing specific HTML files inside the extension package as being sandboxed. Whenever a sandboxed page is loaded, it will be moved to a unique origin, and will be denied access to chrome.* APIs. If we load this sandboxed page into our extension via an iframe, we can pass it messages, let it act upon those messages in some way, and wait for it to pass us back a result. This simple messaging mechanism gives us everything we need to safely include eval-driven code in our extension's workflow.

Creating and using a sandbox.

If you'd like to dive straight into code, please grab the sandboxing sample extension and take off. It's a working example of a tiny messaging API built on top of the Handlebars templating library, and it should give you everything you need to get going. For those of you who'd like a little more explanation, let's walk through that sample together here.

List files in manifest

Each file that ought to be run inside a sandbox must be listed in the extension manifest by adding a sandbox property. This is a critical step, and it's easy to forget, so please double check that your sandboxed file is listed in the manifest. In this sample, we're sandboxing the file cleverly named "sandbox.html". The manifest entry looks like this:

{
  ...,
  "sandbox": {
     "pages": ["sandbox.html"]
  },
  ...
}

Load the sandboxed file

In order to do something interesting with the sandboxed file, we need to load it in a context where it can be addressed by the extension's code. Here, sandbox.html has been loaded into the extension's Event Page (eventpage.html) via an iframe. eventpage.js contains code that sends a message into the sandbox whenever the browser action is clicked by finding the iframe on the page, and executing the postMessage method on its contentWindow. The message is an object containing two properties: context and command. We'll dive into both in a moment.

chrome.browserAction.onClicked.addListener(function() {
 var iframe = document.getElementById('theFrame');
 var message = {
   command: 'render',
   context: {thing: 'world'}
 };
 iframe.contentWindow.postMessage(message, '*');
});

For general information about the postMessage API, take a look at the postMessage documentation on MDN . It's quite complete and worth reading. In particular, note that data can only be passed back and forth if it's serializable. Functions, for instance, are not.

Do something dangerous

When sandbox.html is loaded, it loads the Handlebars library, and creates and compiles an inline template in the way Handlebars suggests:

<script src="handlebars-1.0.0.beta.6.js"></script>
 <script id="hello-world-template" type="text/x-handlebars-template">
   <div class="entry">
     <h1>Hello, {{thing}}!</h1>
   </div>
 </script>
 <script>
   var templates = [];
   var source = document.getElementById('hello-world-template').innerHTML;
   templates['hello'] = Handlebars.compile(source);
 </script>

This doesn't fail! Even though Handlebars.compile ends up using new Function, things work exactly as expected, and we end up with a compiled template in templates[‘hello'].

Pass the result back

We'll make this template available for use by setting up a message listener that accepts commands from the Event Page. We'll use the command passed in to determine what ought to be done (you could imagine doing more than simply rendering; perhaps creating templates? Perhaps managing them in some way?), and the context will be passed into the template directly for rendering. The rendered HTML will be passed back to the Event Page so the extension can do something useful with it later on:

window.addEventListener('message', function(event) {
  var command = event.data.command;
  var name = event.data.name || 'hello';
  switch(command) {
    case 'render':
      event.source.postMessage({
        name: name,
        html: templates[name](event.data.context)
      }, event.origin);
      break;

    // case 'somethingElse':
    //   ...
  }
});

Back in the Event Page, we'll receive this message, and do something interesting with the html data we've been passed. In this case, we'll just echo it out via a Desktop Notification, but it's entirely possible to use this HTML safely as part of the extension's UI. Inserting it via innerHTML doesn't pose a significant security risk, as even a complete compromise of the sandboxed code through some clever attack would be unable to inject dangerous script or plugin content into the high-permission extension context.

This mechanism makes templating straightforward, but it of course isn't limited to templating. Any code that doesn't work out of the box under a strict Content Security Policy can be sandboxed; in fact, it's often useful to sandbox components of your extensions that would run correctly in order to restrict each piece of your program to the smallest set of privileges necessary for it to properly execute. The Writing Secure Web Apps and Chrome Extensions presentation from Google I/O 2012 gives some good examples of these technique in action, and is worth 56 minutes of your time.