tl;dr: ISPs, please reduce your cookie scope.

Everyone now knows that hosting user generated content on a sub-domain is bad. Attacks have been demonstrated on sites such as GitHub, and it’s why Google uses googleusercontent.com.

But what if you’re an ISP. You might not host any user-content, however, you probably assign customers an IP which has Reverse DNS set. You’ll probably see hostnames like 1-1-1-1-ip-static.hfc.comcastbusiness.net or 2.2.2.2.threembb.co.uk.

This isn’t really an issue. The issue is when the hostname assigned is a sub-domain of your own site. If you do this, along with cookies with loose domain scope (fairly common practice), and forward DNS (again, fairly common), then this combination can result in cookie stealing, and therefore account hijacking.

To pull this off, an attacker either needs to be a customer of the ISP they’re targeting, or have access to a machine of a customer (pretty easy with the use of botnets). A web server is then hosted on the connection, and referenced by the hostname assigned (as opposed to the IP).

Example

Rather than showing a real world example, I’d rather keep the companies names private, I’ve setup a proof-of-concept.

We have a fake ISP hosted on fin1te-dsl.com, which mimics an ISPs portal. Registering an account and logging in generates a session cookie (try it out).

We also have a site (152-151-64-212.cust.dsl.fin1te-dsl.com) which in real life would be hosted on a users own connection. A page, 152-151-64-212.cust.dsl.fin1te-dsl.com/debug.php, is hosted to display the cookies back for debug purposes.

Now, we just need a user who has a session to submit a request to our own site and we can grab them. Since we’re accessing the cookies via the HTTP request and not via Javascript, we can write a quick stealer which sets a content-type of image/jpeg and embed the image on a page.

And the cookies show up in the logs.

We just need to set our own cookie to this value and we’ve successfuly hijacked their session.

Out of the four major UK ISPs I tested, two were vulnerable (now patched). If you assume an equal market share (based on 2012 estimates), that’s approximately 10.5 million users who can be potentially targeted. Of course, they have to be logged in - but you can always embed the cookie stealer as an image on a support forum, for example.

Mitigation Techniques

We have three mitigation options. The first is to remove super cookies and restrict the scope to a single domain. This may be impractical if you separate content onto different sub-domains. The second is to disable forward DNS for customers. And the third is to change the hostname assigned to one which isn’t a sub-domain.

In addition, techniques such as pinning a session to an IP address will help to an extent. Unless you store a CSRF token in a cookie, in which case, we can just CSRF the user.

Source

If you want to browse the source code of the proof-of-concept, it’s available on Github.

Note

Since I didn’t have the time to test every single ISP in the world (just the UK ones) for the three requirements that make them vulnerable, I decided to send an email to [email protected] addresses at the top 25 ISPs - 20 of these bounced, and I received no reply from the other 5.

The two UK ones I originally contacted patched promptly and gave good updates, so kudos to you two.

Back in April I found three CSRF issues on Instagram, stemming from their Android/iOS App API (which is slightly different from their public API - it’s hosted on their main domain and doesn’t need an access token).

These issues were present in the following end-points:

  • accounts/remove_profile_pic - This is used to remove the profile picture from an account
  • accounts/set_private - This is used to mark a profile as private
  • accounts/set_public - This is used to mark a profile as public

Obviously the best one out of these is accounts/set_public. With a simple GET request we can reveal anyones profile and access their private pictures. Pretty cool.

Facebook patched the holes pretty quickly and I was awarded a decent bounty for it.

Once patched I checked to make sure that it was indeed fixed, and issuing a GET request returns a 405 Method Not Allowed response.

Round Two

I didn’t blog about the issue and completely forgot about it until recently. I decided to have another look at the Android App to see if there was any new end-points to play around with.

Pretty much all API requests within the app call a method named setSignedBody. This generates a hash of the parameters with a secret embedded in an .so file, meaning we can’t craft our own request on-the-fly and submit on the users behalf (without extracting the secret).

However, the three end-points I submitted still didn’t use setSignedBody (presumably because there are no parameters needed), and therefore no token is sent along. Because of this, we can submit a POST request and still perform the attack which was supposed to be fixed!

The use of setSignedBody without a CSRF token means that all end-points are vulnerable to a replay attack. You simply submit the request yourself, catch the request in Burp, and replay to the victim. Unfortunately, this is something I realised after the bug was fixed, so no screenshots available.

So the moral here is that you should double-double-check that an issue is fixed. If I’d been more thorough in testing the fix I would have spotted it sooner than four months, my bad.

Fix

This is now patched by requiring all requests to have a csrftoken parameter. Any request which is signed also requires a _uid parameter to prevent replay attacks (unless you extract the secret…).

The original proof-of-concept now returns a 400 error.

The response body is a JSON object showing the error message.

I’ve found a few bugs on various Facebook satellite/marketing domains (ones which are part of the Facebook brand, but not necessarily hosted/developed by them, and not under the *.facebook.com domain). Most of them aren’t that serious.

This one isn’t an exception, and I wouldn’t normally blog about it, but it’s an interesting use case as to why content types are important.

The bug is an XSS discovered on Facebook Studio. This is linked to by some Facebook marketing pages, and is used to showcase advertising campaigns on Facebook.

There is an area which allows you to submit work to the Gallery. This form conveniently has an option to scrape details from your Facebook page and fill in boxes for you (such as Company Name, Description).

This calls an AJAX end-point with your pages URL as a parameter.

If we set our pages description to something containing HTML/Javascript, it’s properly escaped. However, it’s escaped client-side. The end-point incorrectly sends a content-type header of text/html, when the response is actually JSON.

When browsed to directly (it doesn’t need any CSRF tokens to be viewed, despite the hash param), we see our script executed.

The cool thing about this bug is that whilst it’s not persistent (the payload is fetched when the page is visited), the code is not present in the request body, therefore avoiding Chrome’s XSS Auditor and IE’s XSS Filter.

Had the content type been set to application/json, the code would have not run (until you start to consider content sniffing…).

Fix

The content type is now set correctly.

Timeline

  • 15th August 2013 - Issue Reported
  • 21st August 2013 - Acknowledgment of Report
  • 21st August 2013 - Issue Fixed

This is a quick post about a simple bug I found on Friendship Pages on Facebook. (Note: Not nearly as cool as a full account takeover, however!)

Friendship Pages show you how two users on Facebook are connected, with posts and photos they’re both tagged in, events they’ve both attended and common friends. On these pages, you’re given the option to upload a cover photo (like you would on your profile, or an event).

Removing A Cover

The cover photo on someones friendship page, we can remove from any account.

First, we need the friendship_id, which can be obtained with an AJAX call to /ajax/timeline/friendship_cover/selector, where profile_id is one user and friend_id is another.

Using this friendship_id we make an AJAX call to /ajax/timeline/friendship_cover/remove, placing the value into the profile_id parameter.

Refresh the page, and it’s disappeared.

Fix

Now, you can only remove your own cover.

Timeline

  • 29th August 2013 - Reported
  • 2nd September 2013 - Acknowledgment of Report
  • 2nd September 2013 - Issue Fixed

This post will demonstrate a simple bug which will lead to a full takeover of any Facebook account, with no user interaction.

Facebook gives you the option of linking your mobile number with your account. This allows you to receive updates via SMS, and also means you can login using the number rather than your email address.

The flaw lies in the /ajax/settings/mobile/confirm_phone.php end-point. This takes various parameters, but the two main are code, which is the verification code received via your mobile, and profile_id, which is the account to link the number to.

The thing is, profile_id is set to your account (obviously), but changing it to your target’s doesn’t trigger an error.

To exploit this bug, we first send the letter F to 32665, which is Facebook’s SMS shortcode in the UK. We receive an 8 character verification code back.

We enter this code into the activation box (located here), and modify the profile_id element inside the fbMobileConfirmationForm form.

Submitting the request returns a 200. You can see the value of __user (which is sent with all AJAX requests) is different from the profile_id we modified.

Note: You may have to reauth after submitting the request, but the password required is yours, not the targets.

An SMS is then received with confirmation.

Now we can initate a password reset request against the user and get the code via SMS.

Another SMS is received with the reset code.

We enter this code into the form, choose a new password, and we’re done. The account is ours.

Fix

Facebook responsed by verifying that you have permission to modify the phone number on the profile denoted by profile_id.

Timeline

  • 23rd May 2013 - Reported
  • 28th May 2013 - Acknowledgment of Report
  • 28th May 2013 - Issue Fixed

Note

The bounty assigned to this bug was $20,000, clearly demonstrating the severity of the issue.