Table of contents

  1. Preface: Giving first insight into the idea and why I think that hooking into a browser is a good idea.
  2. Many different ways to make browser game bots: Discussion various techniques to write HTTP/WebSocket bots
  3. How does chess.com internally look like?: Investigation of the client side behavior of chess.com
  4. How the bot works: Explaining how my shared library hooks firefox network functions
  5. Conclusion: Summary of my discoveries
  6. Demo Video and another, better demo video: You might only watch that video, but make sure you read the explanation on the very bottom of this blog post!
  7. You may find the sources to the shared library (so) on my github account.

Preface

Usually I don't have good ideas in forms of flashes of genius. On the contrary, I think that many endeavors and interesting projects might be reasonable if realized, but often so, there's a huge amount of work involved and too many variables and strategic decisions in the process that could eventually render the project a failure. What I try to say: A mediocre idea well engineered might be a good product. But a good idea badly implemented and designed is usually just bad in it's final form.

Raw, abstract ideas exist in abundance. So does the will to make them happen. But once in a while, I manage to overcome these two hindrances and begin to build and shape my ideas into reality.

So around 10 days ago (yes it took me that long to make it) I went jogging in the forest and my subconscious searched for different possibilities to cheat in a chess browser game (chess.com to be specific). I knew that there were quite some different approaches. So before I get all to technical, let's define the obstacle:

I want build a bot that plays automatic actions/moves on a online chess server, whose communication is realized over HTTP 1/1 and WebSocket (https and wws respectively), all TLS/SSL encrypted.

Additionally the following constraints need to be considered:

  • The bot should work while the player is actually playing. That means: Without any real action by a human, the bot won't do anything at all.
  • The engine that computes the moves is a local process. I chose Stockfish: it has a pretty decent strength and is open source which makes it usable on Linux contrary to Houdini. If we want to use engine calculated information in the browser we need to get a channel to a local process. We can't integrate Stockfish originated data into the browser process space with memory injection methods, since it seems hard and I don't know how to do it (correctly). But we surely may interact with the local process in any other possible IPC way.

So this was my idea that overcame me in the woods while running: Why should I even bother rebuilding the whole chess.com communication protocol, when I can just replace the moves I made ingame with engine calculated equivalents?!

Every move (like a2a3 or g7g8) has identical lengths, so I need to find the point where the packets (or messages in WebSockets RFC terminology) are still decrypted (note that all traffic is TLS/SSL layered) such that it's possible to update them, without dealing with the hassle of TLS/SSL.

And there is only one place where I can do this: Inside of the browser process space, in some networking function that sends and receives the target packets. This was my idea. I knew that it must be possible to find this hook point and to alter the packets somewhere.

But hell, I didn't knew that this approach was that hard to achieve!

So before I'll dive into the technical internals, I'll discuss the different possible architectures and designs of cheating in particular for a chess.com bot, but which are also perfectly applicable to general browser game automation.

Many different ways to make browser game bots!

  1. A high level approach would be to inject custom javascript into the DOM of a running browser session.
    This piece of javascript then connects to a local server (e.g. to a simple server listening on localhost:SOMEPORT) that responses with a bot calculated move when requesting with the current game action history. A pretty nice example of this technique is this video on youtube (I suppose it uses this design paradigm, but I don't know because there are no open sources to investigate further).
    Advantages: Comparably easy implementation, because we don't need to tinker with packets and can directly attack the game logic. Works for all platforms and all browsers, because javascript is universally runnable.
    Drawbacks: Maybe we can't build a connection to localhost due to the same-origin policy (Can we?) This would render this approach impossible, since we absolutely need a way to communicate with the local program that supplies us with engine moves. Then maybe a socket connection (although listening on localhost) might be slower than other IPC techniques, because of the TCP/IP stack overhead.
  2. Another high level technique would be to write a custom browser extension that intercepts the browser game traffic and injects bot calculated moves.
    Advantages: We don't need to hassle with SSL/TLS decryption because our application is executed in the browser process. Furthermore it must be somehow possible to communicate with a local process that supplies the extension with engine knowledge.
    Disadvantages: Platform and browser dependent code. When going down that road, I would need to write platform independent C code for different browsers. Additionally, it's questionable if we can speak to different processes, because of the browser sandboxes and their security policies.
  3. Another approach: Low level network stack interception technique: I could sniff on the network interface where the TCP packets of the communication sessions are exchanged with a appropriate API like scapy or libpcap (When writing C directly). Because the communication is encrypted with SSL/TLS, there needs to be a reliable way to decrpyt the TCP packets. This is not a trivial task, since different browsers use different ciphers/HMACS for TSL/SSL (they are determined in the handshake).
    This issue however is alleviated by the fact that some browsers can be started with a option (environment variable SSLKEYLOGFILE) to dump the current key secrets of a SSL/TLS session to a file which my bot could use to decrypt/encrypt the traffic on the fly.
    Once decrypted, I would modify the appropriate move values and replace the action I made, with the bot calculated move. Then I need to encrypt the packet again with the cipher specified in the SSLKEYLOGFILE. But this implies several issues as for example:
    How can my sniffer/bot know which cipher was used? He only knows the current keys for the session, but not the ciphers. In the worst case, we need to also sniff the SSL/TLS handshake and connection start. This again is a very tedious process since the many quirks and combinations of a TLS session beginning.
    Padding shouldn't be a bigger issue because the moves in the packet are just replaced and no additional data is injected or deleted. So the TCP packets should stay the same. We only need to recalculate the checksums.
  4. Another thoroughly low level approach: Hooking .so libraries! This was my idea while running in the forest!
    This might be the hardest approach in the investigation and research phase (It's not a easy task to find the correct hooking points in the HUGE chromium or firefox code base, but it looks like the NSS library is a good starting point), but also the most elegant way, because we don't need to hassle with SSL/TLS quirks and can work with plaintext HTTP requestes/responses, when we find a good hook point. This approach, although very elegant and straightforward comes with huge drawbacks:
    Hooking is very platform depended. On linux we could hook with kernel modules or the LD_PRELOAD trick. On Windows we'd need other hooking techniques like IAT hooking, using the microsoft hooking library or something else. Additionally, we need to know the target architecutre (Differences between x86 and amd64). But from what I know now, this approach seems to be the best compared to the others, because we can work in the target process space (The browser userland process space itself) and we do not reassemble/decrypt packets ourselves (like in the previous technique). Form grabbers are a related technique used by blackhats for their trojans.
  5. The traditional way: Re-implement the high level browser game protocol by forging messages and HTTP Requests. This is not very elegant because our goal is to just modify packets. But when following this approach, we'd need to rebuilt the whole game logic (or at least many parts). This'd look something like the following:

    import requests
    
    login = requests.post('https://browsergame.com', {'login': user, 'password': pass})
    if login.read().success():
        requeset = request.get('https://browsergame.com/action.php?gamelogic=blabla')
        # Do implement game logic
    
  6. The easy way: Make a click bot! Just control the mouse programmatically and initiate actions like a human would.
    Normally such bots learn where the chess board in the browser window is located in the browser window by comparing the pixel colors of the screen with the color of the chess squares (or whatever field you want to locate). Slightly more adanced would be to locate the html elements that represent the squares and recalculate the exact location based on viewport sizes and settings and rendering. But I don't know how it is actually made, do you have a idea?
    Anyways, the huge majority of chess.com bots is of this architecture type. You can view them in action here or here on youtube.
    Advantages: Very easy to implement, very powerful technique, very good possibilities to evade anti cheating techniques (For instance, if there were cheating detection besides the ones who detect and doom very strong play as engine originated, these approaches would try to find out whether the current player is human by interpreting mouse movements and intervals between moves. But somewhere this information has to be gathered and sent to a server. We could just remove this submittal and therefore the server has no way to learn that we are cheaters).

How does chess.com internals look like?

Having discussed the different architectures for a bot (Remember: I chose the hooking technique!) it was time to learn more about the target.

My plan was very ambitions, because I had very little experience with hooking processes and hell, I didn't code in C or assembly for over 1.5 years, nor did I have much experience in the past. Muppets! Even now, I know nothing about low level stuff at all.

So it was definitely time to get my learning hat on and to investigate how chess.com works.

I based a lot of my investigations on the V8 debugger, the one integrated in the Chrome browser dev tools. I am not much of a Google fan, but Jesus fucking Christ, Chrome's Development tools (the stuff that pops up when you click on Inspect Element in the context menu) is just a very well crafted tool!

It's very fast, newer hangs or remarkably loads and there are so many features whilst maintaining a very clear and concise interface. If you don't know how to use it, I really suggest reading and spending a hour on this tutorial (Do the assignments, otherwise what you learned won't stick): DevTools javascript debugger tutorial.

So by using this debugger I found out that chess.com inner client side machinery essentially consists of around 30k lines of javascript code. There are three major files, LiveChessAdvanced.js, LiveChessBase.js, LiveChessCore.js, you can see a excerpt in the screenshot below:

[caption id="attachment_654" align="alignleft" width="640"]Look into
the chess.com internals with the Chrome debugging
tools Look into the chess.com internals with the Chrome debugging tools[/caption]

Furthermore all of chess.com's networking is handled by a file called cometd.js, so chess.com makes use of the Bayeux protocol. You can learn more about the javascript library on cometd.org.

But why such a strange protocol?

HTTP 1.1 is a request response protocol, it's not suited to implement complex game protocols.

This was quite a issue in the last years, because companies wanted to bring complex applications (like browser games) to the browser and so they implemented a range of techniques that circumvented these obstacles by exploiting some HTTP quirks, like the fact that HTTP connections stay open under specific circumstances, like loading javascript in chunks (A form of long polling, it's basically just creating script elements in excess and pointing their src attribute to the cometd server which then sends back JavaScript (or JSONP) with events as its payload). Or you can achieve alternatively the same with frames. All these techniques are summarized under the umbrella term cometd, and you should read about them on wikipedia.

But in late 2011 the WebSocket standard was established in it's final form and nowadays we don't need to make use of cometd techniques (At least I assume that WebSocket should replace these techniques), because we can make bi-directional, full-duplex (simultaneous communication in both directions) TCP like connections using WebSockets. And this is exactly what chess.com does, it exclusively uses WebSockets for their protocol, when the connection is good enough.

But now comes the first quirk: I strongly assume that this cometd.js library provides fallbacks for the case they detect that WebSockets aren't working for some reason. This happend to me quite often:

My internet connection here is very messy and I have a speed limit of around 60kps downstream and around 10kps upstream. I don't know if this could be a reason, that the library sometimes changes protocols, but it certainly does. I first realized this, when I was able to intercept the game traffic using the HttpHeaders firefox plugin (that, like the name suggests, only captures HTTP traffic) and then suddenly in the next game the captured traffic was empty. So I concluded that the protocol responsible for the game traffic must be chosen dynamically, for which the connection stability seems to be a variable!

So enough written, show me some protocol examples of chess.com you shout! Here you go:

This is a typical WebSocket JSON message from the server that updates the game state of a game (This packet belongs to a game that I was received while kibitzing, not one I played myself):

[{"data":{"sid":"gserv","game":{"id":709324515,"gametype":"chess","time":{"basetime":600,"timeinc":0},"players":[{"uid":"DAHUANXIONG","status":"online","lag":2,"lagms":220,"lightning":2105,"blitz":2064,"standard":1361,"lightning960":1200,"blitz960":1200,"standard960":1200,"bughouse":1200,"ml":10,"title":null,"mod":false,"new":false,"country":"CN","av":true,"avatar":"//d1lalstwiwz2br.cloudfront.net/images_users/avatars/DAHUANXIONG_small.1.jpeg","clientfeatures":{"examineboard":true,"multiplegames":true,"clientname":"LC4Full v2013121801; Chrome/25; Windows"},"nonverified":true},{"uid":"Mrsinj","status":"playing","lag":2,"lagms":259,"gid":709324515,"lightning":2153,"blitz":2205,"standard":1200,"lightning960":1200,"blitz960":1200,"standard960":1200,"bughouse":1200,"ml":10,"title":null,"mod":false,"new":false,"country":"RS","av":true,"avatar":"//d1lalstwiwz2br.cloudfront.net/images_users/avatars/Mrsinj_s.1.jpg","clientfeatures":{"examineboard":false,"multiplegames":true,"clientname":"LC4Simple v2013121801; Chrome/32; Windows"}}]},"tid":"Game"},"channel":"/game/~1"}]

If you're not a fan of such unformatted, jammed JSON data, you should inspect the code on a online JSON editor like www.jsoneditoronline.org.

A few interesting fields: uid identifies the players, status can be starting, in_progress, aborted or stopped, lag and lagms specify the lag of the player (How can you measure that?). Funny enough, the protocol architect uses redundant information here, lag is just the rounded version of lagms, but it doesn't bring other information.

gid is the global game id. For instance, you should be able to revisit the upper game under the following link http://www.chess.com/livechess/game?id=709324515. But this particular message above is not of big interest, because it doesn't contain any moves, it just provides some status information. The packets that are sent on every move made look like the next message:

[{"data":{"sid":"gserv","game":{"id":709324427,"status":"in_progress","seq":9,"players":[{"uid":"ImaginaryDunker","status":"playing","lag":1,"lagms":132,"gid":709324427},{"uid":"Alexander_Donchenko","status":"playing","lag":1,"lagms":115,"gid":709324427}],"moves":"gvYIlBIBvB0KBrZJow","clocks":[582,571],"draws":[],"squares":[0,0]},"tid":"GameState"},"channel":"/game/709324427"}]

Three fields are of particular importance, "moves":"gvYIlBIBvB0KBrZJow", "clocks":[582,571] and "status":"in_progress". They are pretty much self explanatory at this point I suppose.

So let's finally look at a typical game session in terms of parsed WebSocket messages. Please note that I might not include all messages that are exchanged in a game session, simply because I don't care to much for trivial information (like polling the server for specific features). The bottom messages are the ones which count when you want to understand a typical game session:

First of all, a connection message to the server is sent when the browser connects to the server:

1. Initial connection outbound message

[{"channel":"/meta/connect","connectionType":"ssl-websocket","advice":{"timeout":0},"id":"10","clientId":"6djua267g6n8ydja21m8yhztj3gtpx","ext":{"ack":-1,"timesync":{"tc":1390715596738,"l":661,"o":-17527952}}}]

I am not sure, that encoding/encryption the clientId is, but it didn't matter for my purpose to write a bot. So when I expect a challenge from another player, I receive such a packet:

2. Game begins

[{"data":{"sid":"gserv","game":{"id":709341692,"status":"starting","seq":0,"players":[{"uid":"ponkhan","status":"playing","lag":2,"lagms":285,"gid":709341692},{"uid":"EatingSpiders","status":"playing","lag":1,"lagms":177,"gid":709341692}],"abortable":[true,true],"moves":"","clocks":[600,600],"draws":[],"repeated":true,"squares":[0,0]},"tid":"GameState"},"channel":"/game/709341692"}]

Noteworthy is the status "starting". As you can see, the moves value is still empty. Then my adversary made a move and I receive again a message in the style as above:

3. First move received

[{"data":{"sid":"gserv","game":{"id":709341692,"status":"starting","seq":1,"players":[{"uid":"ponkhan","status":"playing","lag":3,"lagms":381,"gid":709341692},{"uid":"EatingSpiders","status":"playing","lag":1,"lagms":183,"gid":709341692}],"abortable":[true,true],"moves":"mC","clocks":[600,600],"draws":[],"squares":[0,0]},"tid":"GameState"},"channel":"/game/709341692"}]

Now moves isn't empty anymore, it contains "mC", but what the heck does it mean?

It's basically a encoding of chess.com for their move notation. I don't know why the chose it, maybe it's represents a compression because it'd use n/2 of the space where n is the number of bytes used in the algebraic chess notation..

But I honestly doubt it, because there are far better compression methods for 64 possible squares (and some promotion combinations).

Maybe it's a way to stop stupid cheaters like me to pursue their devilish deeds, but that can't be possible, since the method that decoded/decrypts the move notation isn't more than a simple look-up table:

Here is the relevant code part of the obfuscated javascript source in LiveChessCore.js, at line 8577 (after applying Chrome's pretty print functionality):

(function() {
    zf9bd6 = function() {
        this.z6439d = {"a1": "a","a2": "i","a3": "q","a4": "y","a5": "G","a6": "O","a7": "W","a8": "4","b1": "b","b2": "j","b3": "r","b4": "z","b5": "H","b6": "P","b7": "X","b8": "5","c1": "c","c2": "k","c3": "s","c4": "A","c5": "I","c6": "Q","c7": "Y","c8": "6","d1": "d","d2": "l","d3": "t","d4": "B","d5": "J","d6": "R","d7": "Z","d8": "7","e1": "e","e2": "m","e3": "u","e4": "C","e5": "K","e6": "S","e7": "0","e8": "8","f1": "f","f2": "n","f3": "v","f4": "D","f5": "L","f6": "T","f7": "1","f8": "9","g1": "g","g2": "o","g3": "w","g4": "E","g5": "M","g6": "U","g7": "2","g8": "!","h1": "h","h2": "p","h3": "x","h4": "F","h5": "N","h6": "V","h7": "3","h8": "?"};
        this.ze8500 = {"_a": "a1","_i": "a2","_q": "a3","_y": "a4","_G": "a5","_O": "a6","_W": "a7","_4": "a8","_b": "b1","_j": "b2","_r": "b3","_z": "b4","_H": "b5","_P": "b6","_X": "b7","_5": "b8","_c": "c1","_k": "c2","_s": "c3","_A": "c4","_I": "c5","_Q": "c6","_Y": "c7","_6": "c8","_d": "d1","_l": "d2","_t": "d3","_B": "d4","_J": "d5","_R": "d6","_Z": "d7","_7": "d8","_e": "e1","_m": "e2","_u": "e3","_C": "e4","_K": "e5","_S": "e6","_0": "e7","_8": "e8","_f": "f1","_n": "f2","_v": "f3","_D": "f4","_L": "f5","_T": "f6","_1": "f7","_9": "f8","_g": "g1","_o": "g2","_w": "g3","_E": "g4","_M": "g5","_U": "g6","_2": "g7","_!": "g8","_h": "h1","_p": "h2","_x": "h3","_F": "h4","_N": "h5","_V": "h6","_3": "h7","_?": "h8"};
        this.z56f6c = "{";
        this.z87e36 = "~";
        this.zbba2c = "}";
        this.z7fd23 = "(";
        this.za93ff = "^";
        this.z536e1 = ")";
        this.z817b2 = "[";
        this.z00e3d = "_";
        this.ze4b5e = "]";
        this.z91a79 = "@";
        this.zc21d8 = "#";
        this.zaa236 = "$";
    };
    zf9bd6.prototype = {z0cd66: function(ze2e37, zbd29b, z2ba00) {
            var zed3d4 = this.z6439d[ze2e37];
            var z98075 = this.z6439d[zbd29b];
            if (z2ba00) {
                var dir = zbd29b.charCodeAt(0) - ze2e37.charCodeAt(0);
                if (z2ba00 == "q")
                    z98075 = (dir == -1) ? this.z56f6c : ((dir == +1) ? this.zbba2c : this.z87e36);
                else 
                if (z2ba00 == "n")
                    z98075 = (dir == -1) ? this.z7fd23 : ((dir == +1) ? this.z536e1 : this.za93ff);
                else 
                if (z2ba00 == "r")
                    z98075 = (dir == -1) ? this.z817b2 : ((dir == +1) ? this.ze4b5e : this.z00e3d);
                else 
                if (z2ba00 == "b")
                    z98075 = (dir == -1) ? this.z91a79 : ((dir == +1) ? this.zaa236 : this.zc21d8);
            }
            return zed3d4 + z98075;
        },zafc74: function(z32182) {
            var zed3d4 = z32182.charAt(0);
            var z98075 = z32182.charAt(1);
            var ze2e37 = this.ze8500["_" + zed3d4];
            var zbd29b;
            var z2ba00 = null;
            var z2b70d = ze2e37.charCodeAt(0);
            var z5b3be = null;
            if (z98075 == this.z56f6c) {
                z5b3be = z2b70d - 1;
                z2ba00 = "q";
            } else 
            if (z98075 == this.z87e36) {
                z5b3be = z2b70d;
                z2ba00 = "q";
            } else 
            if (z98075 == this.zbba2c) {
                z5b3be = z2b70d + 1;
                z2ba00 = "q";
            } else 
            if (z98075 == this.z7fd23) {
                z5b3be = z2b70d - 1;
                z2ba00 = "n";
            } else 
            if (z98075 == this.za93ff) {
                z5b3be = z2b70d;
                z2ba00 = "n";
            } else 
            if (z98075 == this.z536e1) {
                z5b3be = z2b70d + 1;
                z2ba00 = "n";
            } else 
            if (z98075 == this.z817b2) {
                z5b3be = z2b70d - 1;
                z2ba00 = "r";
            } else 
            if (z98075 == this.z00e3d) {
                z5b3be = z2b70d;
                z2ba00 = "r";
            } else 
            if (z98075 == this.ze4b5e) {
                z5b3be = z2b70d + 1;
                z2ba00 = "r";
            } else 
            if (z98075 == this.z91a79) {
                z5b3be = z2b70d - 1;
                z2ba00 = "b";
            } else 
            if (z98075 == this.zc21d8) {
                z5b3be = z2b70d;
                z2ba00 = "b";
            } else 
            if (z98075 == this.zaa236) {
                z5b3be = z2b70d + 1;
                z2ba00 = "b";
            }
            if (z5b3be != null) {
                var zd2788 = ze2e37.charAt(1);
                if (zd2788 == 7)
                    zbd29b = String.fromCharCode(z5b3be) + "8";
                else 
                if (zd2788 == 2)
                    zbd29b = String.fromCharCode(z5b3be) + "1";
                else
                    return null;
            } else {
                zbd29b = this.ze8500["_" + z98075];
                if (!zbd29b)
                    return null;
            }
            var z3cbd9 = {};
            z3cbd9["fromArea"] = ze2e37;
            z3cbd9["toArea"] = zbd29b;
            z3cbd9["additionalInfo"] = z2ba00;
            return z3cbd9;
        },za902f: function(zcc2ef) {
            return this.z6439d[zcc2ef];
        },z398d0: function(zccfb6) {
            return this.ze8500["_" + zccfb6];
        }};
})();

Before I forget it: Why the hell are you guys from chess.com obfuscating your JS code? I mean it probably makes sense to pack and encrypt game binaries of big games (that are still very fast cracked by talanted crackers), but why blur Javascript? It's just too easy to reverse the logic with existing tools such as DevTools or Firebug, so why even bother?

No with the function above, we can look the move up which reveals "mC" == "e3e5". So we now how to encode/decode the move notation. Let's inspect the next message:

4. Writing my first move (outgoing Message)

[{"channel":"/service/user","data":{"move":{"gid":709341692,"seq":1,"uid":"EatingSpiders","move":"YI","clock":600,"clockms":60000,"squared":false},"sid":"gserv","tid":"Move"},"id":"259","clientId":"6djua267g6n8ydja21m8yhztj3gtpx"}]

Nothing new here, this is just my browser that send the move "YI" to the game server. You can look it up now on your own, since I have identified the decoding mechanism above.

Now the next messages are just essentially like the previous two (with status in_progress in place of the initial starting), always receiving the opponents move and then I sending my move. So we have all knowledge about the protocol that we need, now let's discuss how I finally implemented my bot.

How the bot is implemented

As discussed, I chose to hook into low level networking functions in the browser. I chose firefox as my hooking target, because there already exist quite many examples on how to hook networking functions in firefox.

So the technique is basically known as the LD_PRELOAD trick. You can learn about it on this short stackoverflow.com explanation.

There are other, better ways to hook library functions, I really recommend you to read this wonderful article about redirecting shared functions in ELF binaries. But you should have profound knowledge of the ELF format in order to follow;

First, I tried to hook directly into the specific HTTP and WebSocket functions in the firefox network library called netwerk or necko. But firefox is written in C++ and therefore it's not so simple to use the LD_PRELOAD trick, because the functions/class names are mangled. I desperately tried to hook such C++ code and I also managed to to so in a simple example. It looks something like below, but it couldn't actually log some calls. Maybe because the function was never called. Here I try to hook into the GenerateCredentials() function of the nsHttpBasicAuth class. I assume the function fires when you enter HttpAuth credentials in a browser.

#include 
#include 
#include 
#include

#ifndef _GNU_SOURCE
#define _GNU_SOURCE
#endif

namespace mozilla {
namespace net {

typedef unsigned short char16_t;
typedef unsigned int uint32_t;
typedef unsigned long PRUint32; 
typedef PRUint32 nsresult;
#define NS_IMETHODIMP NS_IMETHODIMP_(nsresult)

class nsIHttpAuthenticator {

};

class nsISupports {

};

class nsHttpBasicAuth : public nsIHttpAuthenticator
{
public:
    nsHttpBasicAuth();
    virtual ~nsHttpBasicAuth();
    unsigned long GenerateCredentials(void *authChannel,
                                     const char *challenge,
                                     bool isProxyAuth,
                                     const char16_t *domain,
                                     const char16_t *user,
                                     const char16_t *password,
                                     nsISupports **sessionState,
                                     nsISupports **continuationState,
                                     uint32_t *aFlags,
                                     char **creds);

};

nsHttpBasicAuth::nsHttpBasicAuth()
{
}

nsHttpBasicAuth::~nsHttpBasicAuth()
{
}

unsigned long
nsHttpBasicAuth::GenerateCredentials(void *authChannel,
                                     const char *challenge,
                                     bool isProxyAuth,
                                     const char16_t *domain,
                                     const char16_t *user,
                                     const char16_t *password,
                                     nsISupports **sessionState,
                                     nsISupports **continuationState,
                                     uint32_t *aFlags,
                                     char **creds) {
    fprintf(stdout, "[+] GenerateCredentials hooked!\n");
    // define the class member pointer tyep
    typedef unsigned long (nsHttpBasicAuth::*hookedMethod)(void *, const char *, bool, const char16_t *, const char16_t *, const char16_t *, nsISupports **, nsISupports **, uint32_t *, char **);
    static hookedMethod origMethod = 0;

    void *tmpPtr = dlsym(RTLD_NEXT, "_ZN7mozilla3net15nsHttpBasicAuth19GenerateCredentialsEPvPKcbPKtS6_S6_PPNS0_11nsISupportsES9_PjPPc");
    if (tmpPtr == NULL) {
        fprintf(stderr, "dlsym() couldn't located the symbol :( \n");
        exit(EXIT_FAILURE);
    }

    memcpy(&origMethod, &tmpPtr, sizeof(&tmpPtr));

    // call the original method
    unsigned long retVal = (this->*origMethod)(authChannel, challenge, isProxyAuth, domain, user, password, sessionState, continuationState, aFlags, creds);

    return retVal;
}

}
}

Anyways, if I could get succeed in hooking C++ functions by overcoming the mangling issues, I would directly hook into functions found in the mozilla source tree /mozilla-central/source/netwerk/protocol/websocket/. It would be enormously juicy to hook into the file WebSocketChannel.cpp. For example the function void WebSocketChannel::BeginOpen() on line 1061 looks like a nice start. Can anyone provide me a hooking example for this function? That'd be awesome!

So after some further research and giving up trying to hook C++ functions, I decided to hook into firefox's PR_write() and PR_read() low level networking functions, implemented in plain C. If you search for these functions in the internet, you will find lot's of different example for form grabbers, techniques that are commonly used by maleware writers.

After experimenting a little bit with these two functions, I figured out that basically the whole freaking internet flows through them. For instance, if you just open a browser without loading any site, a whole bunch of meta protocols are squeezed through PR_write() and PR_read().

I assume that all possible protocols (like accessing cookies, about:blank, file: and the like) are also handled these functions. It's really funny what you're able to see when you hook into these functions. To formulate it pregnant: Stuff that you thought to be deleted after clearing the cache is happily appearing inside these two functions...

Anyways, the basic C code that implements this hooking technique in form of a shared library looks like the bottom code excerpt. You can always visit my public repository, where you can read the current state of the cheat code.

PRInt32 PR_Write(PRFileDesc *fd, const void *buf, PRInt32 amount) {
    static PRInt32 (*my_PR_write)(PRFileDesc *, const void *, PRInt32) = NULL;

    // Detect the client WebSocket connection attempt
    if (!(webSocketState.state & WEB_SOCKET_CONNECTION_REQUESTED) && amount > 600 && amount < 1200 && memContains((const char *)buf, amount, "Upgrade: websocket") 
        && memContains((const char *)buf, amount, "Origin: http://live.chess.com") && memContains((const char *)buf, amount, "Sec-WebSocket-Key:") &&
             memContains((const char *)buf, amount, "Sec-WebSocket-Version: 13")) {
        webSocketState.state |= WEB_SOCKET_CONNECTION_REQUESTED; 
#if DEBUG_LEVEL >= 1
        INFO_PRINT("Chess.com WebSocket client request detected", SH_RED);
#endif
        // Also start up the Stockfish engine
        //INFO_PRINT("Starting the engine...", SH_BLUE);
        //initStockfish();
    }
    /*
     * When the WebSocket is open sniff for move packets that are between 230 and 240 bytes in size.
     * I observed that these synchronize with moves made and hence I assume these packetses transmit moves.
     * Now I just need do decrypt them and good is.
     */
    if ((webSocketState.state & WEB_SOCKET_CONNECTION_REQUESTED) && amount > 210 && amount < 260) {
#if DEBUG_LEVEL >= 2
        INFO_PRINT("PR_Write() called and buffer size suggests that this packet represents a move!", SH_RED);
#endif
        // Inject a engine made move and update the game state
        modifyMove((char *)buf, (size_t)amount);
    }

    my_PR_write = dlsym(RTLD_NEXT, "PR_Write");

    if (my_PR_write == NULL) {
        fprintf(stderr, "dlsym() failed for PR_Write\n");
        exit(EXIT_FAILURE);
    }

    PRInt32 retVal = (*my_PR_write)(fd, buf, amount);

    return retVal;
}

PRInt32 PR_Read(PRFileDesc *fd, void *buf, PRInt32 amount) {
    static PRInt32 (*my_PR_read)(PRFileDesc *, void *, PRInt32);

    /* 
     * There is a hell of a lot data that passes through this function. Hence we only want
     * to consume as few as possible processing power. We are only interested in data that
     * is in a specfic length range, since it seems that the WebSocket packets on live.chess.com
     * use pretty regular packet size.
     */

    /* 
     * Sniff for incoming live.chess.com WebSocket move data.
     */
    if ((webSocketState.state & WEB_SOCKET_CONNECTION_REQUESTED) && amount > 50 && memContains((const char *)buf, amount, "moves") != NULL) {
        couldRead = 1;
        collectGameState((char *)buf, amount);
#if DEBUG_LEVEL >= 2
        INFO_PRINT("PR_Read() called with 'moves' keyword inside", SH_BLUE);
#endif
    }

    my_PR_read = dlsym(RTLD_NEXT, "PR_Read");

    if (my_PR_read == NULL) {
        fprintf(stderr, "dlsym() failed for PR_Write\n");
        exit(EXIT_FAILURE);
    }

    PRInt32 retVal = (*my_PR_read)(fd, buf, amount);

    return retVal;
}

I hope the above code is more or less self explanatory. But in case it's not, here's a short summary of the rough algorithm that it computes:

1. The shared library (let's call it libpwh.so) is dynamically loaded in a firefox process with the LD_PRELOAD trick. For instance: export LD_PRELOAD=$PWD/libpwh.so; /usr/bin/firefox
2. The functions PR_Read() and PR_Write() are hooked.
3. As soon as a WebSocket request that indicates the beginning of a new live.chess.com chess game is detected, the shared library (the above code) initializes the local Stockfish chess engine with the function int initStockfish() {There should be enough time before the game actually begins to pre-calculate some moves}.
3.1 When a JSON message like the following

[{"data":{"sid":"gserv","game":{"id":706548893,"status":"starting","seq":0,"players":[{"uid":"rocchen","status":"playing","lag":4,"lagms":415,"gid":706548893},{"uid":"workcentre7328","status":"playing","lag":2,"lagms":210,"gid":706548893}],"abortable":[true,true],"moves":"","clocks":[600,600],"draws":[],"repeated":true,"squares":[0,0]},"tid":"GameState"},"channel":"/game/706548893"}]

is intercepted, a game has begun. The needles that indicate such a beginning are thus the players UID and the string "status":"starting"
4. From now on, every outgoing packet (from PR_Write()), that causes a move, must be updated with a chess engine move. This is done with with function void modifyMove(), which in turns relies on a correct game state. Such a outbound packet looks like (We have discussed the protocol in the previous chapter):

[{"channel":"/service/user","data":{"move":{"gid":706662190,"seq":11,"uid":"rocchen","move":"9I","clock":76,"clockms":7661,"squared":false},"sid":"gserv","tid":"Move"},"id":"105","clientId":"4gzikii6gu01a1k7kvdcb7a1h2gh6"}]

5. Concurrently to step 4, the move made by the opponent is synchronized with a local gameState struct variable. The opponent's move is obtained in PR_Read() and the function void collectGameState() keeps the move history current. The function collectGameState() also stores the current game state in a C struct:

// This struct holds the current game state and all kids of game information as parsed and extracted of the live session
typedef struct {
    char status[0x50]; // The status as defined by the protocol. Will be most likely 'playing' during a game.
    uint64_t gameID; // The global game id. Can be looked up after games to review the game.
    uint32_t moveNumber; // The number of moves I did
    char playerName[0x50]; // Me
    char opponentPlayerName[0x50]; // Poor opponent
    uint64_t remainingTimeMsSelf; // yeah does that matter?
    uint64_t remainingTimeMsOpponent; // probably more relevant
    uint16_t lagMsSelf; // i have a stupid connection here in my house
    uint16_t lagMsOpponent;
    char movesMade[0x100]; // in live.chess.com encoded version
    char movesMadeDecoded[0x400]; // the decoded version
    char currentMoveOpponent[0x3]; // last move of opponent
    char currentMoveSelf[0x3]; // my last move
    char currentMoveOpponentDecoded[0x6];
    char currentMoveSelfDecoded[0x6];
    char engineMoves[0x100];
    char engineSuggestion[0x6]; // The best move as suggested by the engine
} _CHESS_COM_GAME_SESSION_STATE;

Conclusion

Hooking is a dirty business and it's not the good idea I thought it was. Therefore my idea was rather average and the implementation is pretty bad. All in all, I am still satisfied, because it works!

It would be a probably good approach if I managed to hook directly into WebSocketChannel.cpp, because I wouldn't need to cumbersomly pick my target messages in the huge traffic that flows thourgh PR_Read().

It's a really strange bot to use, because while my bot modifies packets, the javascript frontend can't properly deal with this fact and in displays a huge mess! This is really interesting because the frontend (in flash and javascript) notes that I made a specific move and saves it accordingly in their variables/objects, but suddenly the game server says that I made in reality atotally different move (The one which was injected by my code). So the poor frontend is heavily confused and sometimes needs up to 10 moves to update the real moves into the frontend browser window. You can observe this behaviour in the video. It's really funny!

To be honest, my bot has still many open issues and sometimes the firefox process crashes randomly. Maybe there's a memory leak somewhere, because after several games someone eats up the whole heap and the whole linux system crashes.

I am pretty sure it's not my malloc/calloc, because I strictly checked to free() all allocated buffers. But on the other hand, I must be the culprit, because without me hooking into firefox, everything works like a charm.

Either I am to inexperienced or it just is hard, but I think making a click bot (as discussed above) for browser games it a far better approach, if the case you're strictly result oriented :)

Demonstration videos

These two videos are of slightly low quality, but you can still comfortably recolonize the chess game:

Please note the following to understand the video: I make random moves with the king. They are not actually being submitted to the server, they are just shown so in the client side interface (That what you can see in the browser). The real moves are updated and modified by the Stockfish engine and then injected into the WebSocket messages. That's also the reason sometimes the chessboard suddenly updates, because the client side GUI recognizes that it showed the false moves. Then you can see the real moves for a second again. ou can see the real moves for a second again.