Some time back while I was working on getting the Javascript Windows Store app platform running on Windows Phone (now available on the last Windows Phone release!) I had an interesting bug that in retrospect is amusing.

I had just finished a work item to get accessibility working for JS WinPhone apps when I got a new bug: With some set of JS apps, accessibility appeared to be totally broken. At that time in development the only mechanism we had to test accessibility was a test tool that runs on the PC, connects to the phone, and dumps out the accessibility tree of whatever app is running on the phone. In this bug, the tool would spin for a while and then timeout with an error and no accessibility information.

My first thought was this was an issue in my new accessibility code. However, debugging with breakpoints on my code I could see none of my code was run nor the code that should call it. The code that called that code was a more generic messaging system that hit my breakpoints constantly.

Rather than trying to work backward from the failure point, I decided to try and narrow down the repro and work forwards from there. One thing all the apps with the bug had in common was their usage of WinJS, but not all WinJS apps demonstrated the issue. Using a binary search approach on one such app I removed unrelated app code until all that was left was the app's usage of the WinJS AppBar and the bug still occurred. I replaced the WinJS AppBar usage with direct usage of the underlying AppBar WinRT APIs and continued.

Only some calls to the AppBar WinRT object produced the issue:

        var appBar = Windows.UI.WebUI.Core.WebUICommandBar.getForCurrentView(); 
        // appBar.opacity = 1;
        // appBar.closeDisplayMode = Windows.UI.WebUI.Core.WebUICommandBarClosedDisplayMode.default;
        appBar.backgroundColor = Windows.UI.Colors.white; // Bug! 

Just setting the background color appeared to cause the issue and I didn't even have to display the AppBar. Through additional trial and error I was blown away to discover that some colors I would set caused the issue and other colors did not. Black wouldn't cause the issue but transparent black would. So would aqua but not white.

I eventually realized that predefined WinRT color values like Windows.UI.Colors.aqua would cause the issue while JS literal based colors didn't cause the issue (Windows.UI.Color is a WinRT struct which projects in JS as a JS literal object with the struct members as JS object properties so its easy to write something like {r: 0, g: 0, b: 0, a: 0} to make a color) and I had been mixing both in my tests without realizing there would be a difference. I debugged into the backgroundColor property setter that consumed the WinRT color struct to see what was different between Windows.UI.Colors.black and {a: 1, r: 0, g: 0, b: 0} and found the two structs to be byte wise exactly the same.

On a hunch I tried my test app with only a reference to the color and otherwise no interaction with the AppBar and not doing anything with the actual reference to the color: Windows.UI.Colors.black;. This too caused the issue. I knew that the implementation for these WinRT const values live in a DLL and guessed that something in the code to create these predefined colors was causing the issue. I debugged in and no luck. Now I also have experienced crusty code that would do exciting things in its DllMain, the function that's called when a DLL is loaded into the process so I tried modifying my C++ code to simply LoadLibrary the DLL containing the WinRT color definition, windows.ui.xaml.dll and found the bug still occurred! A short lived moment of relief as the world seemed to make sense again.

Debugging into DllMain nothing interesting happened. There were interesting calls in there to be sure, but all of them behind conditions that were false. I was again stumped. On another hunch I tried renaming the DLL and only LoadLibrary'ing it and the bug went away. I took a different DLL renamed it windows.ui.xaml.dll and tried LoadLibrary'ing that and the bug came back. Just the name of the DLL was causing the issue.

I searched for the DLL name in our source code index and found hits in the accessibility tool. Grinning I opened the source to find that the accessibility tool's phone side service was trying to determine if a process belonged to a XAML app or not because XAML apps had a different accessibility contract. It did this by checking to see if windows.ui.xaml.dll was loaded in the target process.

At this point I got to fix my main issue and open several new bugs for the variety of problems I had just run into. This is a how to on writing software that is difficult to debug.

When writing a JavaScript library that uses postMessage and the message event, I must be considerate of other JS code that will be running along side my library. I shouldn't assume I'm the only sender and receiver on a caller provided MessagePort object. This means obviously I should use addEventListener("message" rather than the onmessage property (see related What if two programs did this?). But considering the actual messages traveling over the message channel I have the issue of accidentally processing another libraries messages and having another library accidentally process my own message. I have a few options for playing nice in this regard:

Require a caller provided unique MessagePort

This solves the problem but puts a lot of work on the caller who may not notice nor follow this requirement.

Uniquely mark my messages

To ensure I'm acting upon my own messages and not messages that happen to have similar properties as my own, I place a 'type' property on my postMessage data with a value of a URN unique to me and my JS library. Usually because its easy I use a UUID URN. There's no way someone will coincidentally produce this same URN. With this I can be sure I'm not processing someone else's messages. Of course there's no way to modify my postMessage data to prevent another library from accidentally processing my messages as their own. I can only hope they take similar steps as this and see that my messages are not their own.

Use caller provided MessagePort only to upgrade to new unique MessagePort

I can also make my own unique MessagePort for which only my library will have the end points. This does still require the caller to provide an initial message channel over which I can communicate my new unique MessagePort which means I still have the problems above. However it clearly reduces the surface area of the problem since I only need once message to communicate the new MessagePort.

The best solution is likely all of the above.
Photo is Sharing by leezie5. Two squirrels sharing food hanging from a bird feeder. Used under Creative Commons license Attribution-NonCommercial-NoDerivs 2.0 Generic.

My first app for Windows 8 was Shout Text. You type into Shout Text, and your text is scaled up as large as possible while still fitting on the screen, as you type. It is the closest thing to a Hello World app as you'll find on the Windows Store that doesn't contain that phrase (by default) and I approached it as the simplest app I could make to learn about Windows modern app development and Windows Store app submission.

I rely on WinJS's default layout to use CSS transforms to scale up the user's text as they type. And they are typing into a simple content editable div.

The app was too simple for me to even consider using ads or charging for it which I learned more about in future apps.

The first interesting issue I ran into was that copying from and then pasting into the content editable div resulted in duplicates of the containing div with copied CSS appearing recursively inside of the content editable div. To fix this I had to catch the paste operation and remove the HTML data from the clipboard to ensure only the plain text data is pasted:

        function onPaste() {
            var text;

            if (window.clipboardData) {
                text = window.clipboardData.getData("Text").toString();
                window.clipboardData.clearData("Html");
                window.clipboardData.setData("Text", util.normalizeContentEditableText(text));
            }
        }
        shoutText.addEventListener("beforepaste", function () { return false; }, false);
        shoutText.addEventListener("paste", onPaste, false);

I additionally found an issue in IE in which applying a CSS transform to a content editable div that has focus doesn't move the screen position of the user input caret - the text is scaled up or down but the caret remains the same size and in the same place on the screen. To fix this I made the following hack to reapply the current cursor position and text selection which resets the screen position of the user input caret.

        function resetCaret() {
            setTimeout(function () {
                var cursorPos = document.selection.createRange().duplicate();
                cursorPos.select();
            }, 200);
        }

        shoutText.attachEvent("onresize", function () { resetCaret(); }, true);

Level 8 of the Stripe CTF is a password server that returns success: true if and only if the password provided matches the password stored directly via a RESTful API and optionally indirectly via a callback URI. The solution is side channel attack like a timing attack but with ports instead of time.

(I found this in my drafts folder and had intended to post a while ago.)

Code

    def nextServerCallback(self, data):
        parsed_data = json.loads(data)
        # Chunk was wrong!
        if not parsed_data['success']:
            # Defend against timing attacks
            remaining_time = self.expectedRemainingTime()
            self.log_info('Going to wait %s seconds before responding' %
                          remaining_time)
            reactor.callLater(remaining_time, self.sendResult, False)
            return

        self.checkNext()

Issue

The password server breaks the target password into four pieces and stores each on a different server. When a password request is sent to the main server it makes requests to the sub-servers for each part of the password request. It does this in series and if any part fails, then it stops midway through. Password requests may also be made with corresponding URI callbacks and after the server decides on the password makes an HTTP request on the provided URI callbacks saying if the password was success: true or false.
A timing attack looks at how long it took for a password to be rejected and longer times could mean a longer prefix of the password was correct allowing for a directed brute force attack. Timing attacks are prevented in this case by code on the password server that attempts to wait the same amount of time, even if the first sub-server responds with false. However, the server uses sequential outgoing port numbers shared between the requests to the sub-servers and the callback URIs. Accordingly, we can examine the port numbers on our callback URIs to direct a brute force attack.
If the password provided is totally incorrect then the password server will contact one sub-server and then your callback URI. So if you see the remote server's port number go up by two when requesting your callback URI, you know the password is totally incorrect. If by three then you know the first fourth of the password is correct and the rest is incorrect. If by four then two fourths of the password is correct. If by five then four sub-servers were contacted so you need to rely on the actual content of the callback URI request of 'success: true' or 'false' since you can't tell from the port change if the password was totally correct or not.
The trick in the real world is false positives. The port numbers are sequential over the system, so if the password server is the only thing making outgoing requests then its port numbers will also be sequential, however other things on the system can interrupt this. This means that the password server could contact three sub-servers and normally you'd see the port number increase by four, but really it could increase by four or more because of other things running on the system. To counteract this I ran in cycles: brute forcing the first fourth of the password and removing any entry that gets a two port increase and keeping all others. Eventually I could remove all but the correct first fourth of the password. And so on for the next parts of the password.
I wrote my app to brute force this in Python. This was my first time writing Python code so it is not pretty.

Level 7 of the Stripe CTF involved running a length extension attack on the level 7 server's custom crypto code.

Code

@app.route('/logs/')
@require_authentication
def logs(id): 
    rows = get_logs(id) 
    return render_template('logs.html', logs=rows) 

...

def verify_signature(user_id, sig, raw_params):
    # get secret token for user_id
    try:
        row = g.db.select_one('users', {'id': user_id})
    except db.NotFound:
        raise BadSignature('no such user_id')
    secret = str(row['secret'])

    h = hashlib.sha1()
    h.update(secret + raw_params)
    print 'computed signature', h.hexdigest(), 'for body', repr(raw_params)
    if h.hexdigest() != sig:
        raise BadSignature('signature does not match')
    return True

Issue

The level 7 web app is a web API in which clients submit signed RESTful requests and some actions are restricted to particular clients. The goal is to view the response to one of the restricted actions. The first issue is that there is a logs path to display the previous requests for a user and although the logs path requires the client to be authenticatd, it doesn't restrict the logs you view to be for the user for which you are authenticated. So you can manually change the number in the '/logs/[#]' to '/logs/1' to view the logs for the user ID 1 who can make restricted requests. The level 7 web app can be exploited with replay attacks but you won't find in the logs any of the restricted requests we need to run for our goal. And we can't just modify the requests because they are signed.

However they are signed using their own custom signing code which can be exploited by a length extension attack. All Merkle–Damgård hash algorithms (which includes MD5, and SHA) have the property that if you hash data of the form (secret + data) where data is known and the length but not content of secret is known you can construct the hash for a new message (secret + data + padding + newdata) where newdata is whatever you like and padding is determined using newdata, data, and the length of secret. You can find a sha-padding.py script on VNSecurity blog that will tell you the new hash and padding per the above. With that I produced my new restricted request based on another user's previous request. The original request was the following.

count=10&lat=37.351&user_id=1&long=%2D119.827&waffle=eggo|sig:8dbd9dfa60ef3964b1ee0785a68760af8658048c

The new request with padding and my new content was the following.

count=10&lat=37.351&user_id=1&long=%2D119.827&waffle=eggo%80%02%28&waffle=liege|sig:8dbd9dfa60ef3964b1ee0785a68760af8658048c

My new data in the new request is able to overwrite the waffle parameter because their parser fills in a map without checking if the parameter existed previously.

Notes

Code review red flags included custom crypto looking code. However I am not a crypto expert and it was difficult for me to find the solution to this level.

Level 5 of the Stripe CTF revolved around a design issue in an OpenID like protocol.

Code

    def authenticated?(body)
      body =~ /[^\w]AUTHENTICATED[^\w]*$/
    end

...

      if authenticated?(body)
        session[:auth_user] = username
        session[:auth_host] = host
        return "Remote server responded with: #{body}." \
               " Authenticated as #{username}@#{host}!"

Issue

This level is an implementation of a federated identity protocol. You give it an endpoint URI and a username and password, it posts the username and password to the endpoint URI, and if the response is 'AUTHENTICATED' then access is allowed. It is easy to be authenticated on a server you control, but this level requires you to authenticate from the server running the level. This level only talks to stripe CTF servers so the first step is to upload a document to the level 2 server containing the text 'AUTHENTICATED' and we can now authenticate on a level 2 server. Notice that the level 5 server will dump out the content of the endpoint URI and that the regexp it uses to detect the text 'AUTHENTICATED' can match on that dump. Accordingly I uploaded an authenticated file to

https://level02-2.stripe-ctf.com/user-ajvivlehdt/uploads/authenticated

Using that as my endpoint URI means authenticating as level 2. I can then choose the following endpoint URI to authenticate as level 5.

https://level05-1.stripe-ctf.com/user-qtoyekwrod/?pingback=https%3A%2F%2Flevel02-2.stripe-ctf.com%2Fuser-ajvivlehdt%2Fuploads%2Fauthenticated&username=a&password=a

Navigating to that URI results in the level 5 server telling me I'm authenticated as level 2 and lists the text of the level 2 file 'AUTHENTICATED'. Feeding this back into the level 5 server as my endpoint URI means level 5 seeing 'AUTHENTICATED' coming back from a level 5 URI.

Notes

I didn't see any particular code review red flags, really the issue here is that the regular expression testing for 'AUTHENTICATED' is too permisive and the protocol itself doesn't do enough. The protocol requires only a set piece of common literal text to be returned which makes it easy for a server to accidentally fall into authenticating. Having the endpoint URI have to return variable text based on the input would make it much harder for a server to accidentally authenticate.

Level 4 and level 6 of the Stripe CTF had solutions around XSS.

Level 4

Code

> Registered Users 

<%@registered_users.each do |user| %>
<%last_active = user[:last_active].strftime('%H:%M:%S UTC') %>
<%if @trusts_me.include?(user[:username]) %>

• 
<%= user[:username] %>
(password: <%= user[:password] %>, last active <%= last_active %>)


Issue

The level 4 web application lets you transfer karma to another user and in doing so you are also forced to expose your password to that user. The main user page displays a list of users who have transfered karma to you along with their password. The password is not HTML encoded so we can inject HTML into that user's browser. For instance, we could create an account with the following HTML as the password which will result in XSS with that HTML:

This HTML runs script that uses jQuery to post to the transfer URI resulting in a transfer of karma from the attacked user to the attacker user, and also the attacked user's password.

Notes

Code review red flags in this case included lack of encoding when using user controlled content to create HTML content, storing passwords in plain text in the database, and displaying passwords generally. By design the web app shows users passwords which is a very bad idea.

Level 6

Code

...

    def self.safe_insert(table, key_values)
      key_values.each do |key, value|
        # Just in case people try to exfiltrate
        # level07-password-holder's password
        if value.kind_of?(String) &&
            (value.include?('"') || value.include?("'"))
          raise "Value has unsafe characters"
        end
      end

      conn[table].insert(key_values)
    end

Issue

This web app does a much better job than the level 4 app with HTML injection. They use encoding whenever creating HTML using user controlled data, however they don't use encoding when injecting JSON data into script (see post_data initialization above). This JSON data is the last five most recent messages sent on the app so we get to inject script directly. However, the system also ensures that no strings we write contains single or double quotes so we can't get out of the string in the JSON data directly. As it turns out, HTML lets you jump out of a script block using no matter where you are in script. For instance, in the middle of a value in some JSON data we can jump out of script. But we still want to run script, so we can jump right back in. So the frame so far for the message we're going to post is the following: