Fuzzing¶

We talk about securing software by program analysis. We discuss about fuzzing techniques, their advantages and disadvantages and about hybrid approaches.

Slides for this session:

slides

Tasks¶

We higly recommend you use teams of 2 or 3 to work on the tutorials.

Setup¶

For the AFL and KLEE tutorial, you require access to a pre-configured environment. These environments use Docker, both for AFL and KLEE.

You can either:

Download a Virtual Machine that is already configured to use Docker to run the AFL and KLEE containers
Install Docker on your local machine and configure the AFL and KLEE containers

Depending on which method you want to use, skip to the matching section below:

Virtual Machine¶

The easiest way is to use the SIS - Fuzzing virtual machine. This virtual machine has Docker installed and it also has the AFL and KLEE containers configured.

Download the virtual machine (in OVA format) using Bittorrent. Use this Bittorrent file.

The virtual machine is rather large (5.4GB the OVA file, 13GB the unpacked virtual machine); be sure you have disk space. Import the OVA file in VMware or VirtualBox.

Once you start the virtual machine, use the username student and password student. Then start and access the AFL and KLEE Docker instances as the following snippet explains:

student@sis-fuzzing:~$ docker ps
CONTAINER ID   IMAGE     COMMAND   CREATED   STATUS    PORTS     NAMES
student@sis-fuzzing:~$ docker ps -a
CONTAINER ID   IMAGE           COMMAND           CREATED          STATUS                        PORTS     NAMES
f2d5777932b3   klee/klee:2.1   "/bin/bash"       8 minutes ago    Exited (0) 5 minutes ago                klee
5275c2d2c1ed   afltraining     "entrypoint.sh"   19 minutes ago   Exited (137) 14 minutes ago             afl
student@sis-fuzzing:~$ docker start afl
afl
student@sis-fuzzing:~$ docker start klee
klee
student@sis-fuzzing:~$ docker ps
CONTAINER ID   IMAGE           COMMAND           CREATED          STATUS         PORTS                   NAMES
f2d5777932b3   klee/klee:2.1   "/bin/bash"       8 minutes ago    Up 4 seconds   0.0.0.0:23000->22/tcp   klee
5275c2d2c1ed   afltraining     "entrypoint.sh"   20 minutes ago   Up 7 seconds   0.0.0.0:22000->22/tcp   afl

student@sis-fuzzing:~$ docker exec -it klee /bin/bash
klee@f2d5777932b3:~$ ls
klee_build  klee_src
klee@f2d5777932b3:~$ exit

student@sis-fuzzing:~$ ssh fuzzer@localhost -p 22000
fuzzer@localhost's password:
Welcome to Ubuntu 20.04.1 LTS (GNU/Linux 4.15.0-130-generic x86_64)
 * Documentation:  https://help.ubuntu.com
 * Management:     https://landscape.canonical.com
 * Support:        https://ubuntu.com/advantage
Last login: Tue Jan 12 09:52:37 2021 from 172.17.0.1
-bash: warning: setlocale: LC_ALL: cannot change locale (en_US.UTF-8)
fuzzer@5275c2d2c1ed:~$ ls
AFLplusplus  workshop
fuzzer@5275c2d2c1ed:~$ logout

Important

The password for the AFL Docker container is afl. No password is required for the KLEE Docker container.

Local Docker Installation¶

Otherwise you can do your own setup using Docker. If that is the case, do the following steps.

(Skip this if Docker is already installed) Install Docker.

sudo apt-get update
sudo apt-get install -y apt-transport-https ca-certificates curl gnupg-agent software-properties-common
curl -fsSL https://download.docker.com/linux/ubuntu/gpg | sudo apt-key add -
sudo add-apt-repository "deb [arch=amd64] https://download.docker.com/linux/ubuntu $(lsb_release -cs) stable"
sudo apt-get update
sudo apt-get install -y docker-ce docker-ce-cli containerd.io docker-compose
sudo service docker start
sudo usermod -aG docker $USER
sudo newgrp docker

You may need to reautenticate the current user.

Install the AFL Docker container, by following the AFL instructions.

docker pull ghcr.io/mykter/fuzz-training:latest
docker run -d --name afl --privileged -p 2222:2222 -e PASSMETHOD=env -e PASS=afl ghcr.io/mykter/fuzz-training
ssh fuzzer@localhost -p 2222 # use password afl
sudo ~/AFLplusplus/afl-system-config

Install the KLEE Docker container, by followint the KLEE instructions.

The important bit is:

docker pull klee/klee:2.3
docker run -d --name klee --rm -ti --ulimit='stack=-1:-1' klee/klee:2.3
docker exec -it klee /bin/bash

Fuzzing with AFL¶

We will go through the AFL tutorial. Just as the README suggests, we will go over the quickstart and harness exercises.

First, log into the AFL Docker container. Then go inside the workshop/ folder. Then go the quickstart/ folder and follow the instructions.

After launching afl-fuzz make sure to take a look at understanding the status screen.

Once you have some crashes being detected, you can move on to the next step.

You can test the crashes in the out/crashes/:

./vulnerabable < out/crashes/<id_of_crash-file>

Create a new folder with inputs and use that.

Now create your own test file, with an obvious issue / vulnerability (such as a buffer overflow). And build that with AFL support and test / fuzz it using AFL.

Now go through the harness/ folder. Follow the instructions.

Go through as many challenges in the challenges/ folder as you can.

Symbolic Execution with KLEE¶

We will go through the KLEE tutorials. Go through the tutorials in order: First tutorial, Second tutorial, Using symbolic environment.

Clone the official repo to have access to the examples/ directory. Use the directory to work on the exercises.

The latest version of the container has clang installed in /usr/lib/llvm-11/bin/clang. This path is not included in the PATH variable. You can either include it manually or call clang with the full path.

Extra: Fuzzing with driller¶

BONUS: Traditional fuzzers are not able to generate inputs that pass complex string comparisons, magic numbers, etc. Even if a grey-box fuzzer like libFuzzer knows the coverage state and drives the test generation to uncovered paths, this approach won’t determine proper value for complex compares. In this situation, a mix system is used.

This approach knows the set of paths where execution can reach only one branch of the comparison during multiple tests. We say that the fuzzer is blocked in certain instruction pointers. To unblock the code discovery, symbolic execution is used.

In this lab we experiment this situation with driller (https://github.com/shellphish/driller). Take a look on how it works in this example (https://github.com/shellphish/driller/blob/master/README.md).

From session task archive, access hard-to-fuzz subfolder and inspect fuzzy using a static analyzer.

Run fuzzy-fuzzer to see whether issues are found in reasonable time.
Write a Driller script to detect the deep-hidden vulnerability. Driller is installed in this virtual machine (https://repository.grid.pub.ro/cs/hexcellents/sss/SSS%20-%20Ubuntu%2014.04.1%2064-bit.ova). To start driller virtual environment, run workon driller. Be patient! Pay attention to how crash-me reads input!