Using S2E to generate PoVs for Linux, Windows, and CGC binaries

In this tutorial, you will learn:

  • How to configure S2E in order to automatically generated proofs of vulnerability (PoVs) for Linux, Windows, and Decree binaries.

  • How S2E implements PoV generation, so that you can extend it for your own needs. We will show an overview of the S2E plugin architecture as well as the plugins involved in PoV generation.


Before starting this tutorial, make sure that…
  • you understand the theory behind automated PoV generation

  • you have a working S2E environment ready (e.g, in the ~/s2e/env/ directory)

Quickstart on Windows and Linux

S2E comes with a vulnerabilities demo which demonstrates several types of vulnerabilities that S2E can detect. You can find the corresponding source file in vulnerabilities.c.

  1. Build the guest tools. We will use the 32-bit Linux version here. You may also use the 64-bit version and/or the Windows version (in guest-tools{32|64}-win).

$ cd ~/s2e/env/build/s2e/guest-tools32
$ make
  1. Create an analysis project

$ cd ~/s2e/env
$ s2e new_project -i debian-12.5-i386 -n vuln-lin32-32 --tools=pov \
  build/s2e/guest-tools32/common/demos/vulnerabilities @@

Here is a breakdown of the s2e new_project command above:

  • It creates a new project called vuln-lin32-32 that will run on the debian-12.5-i386 image.

  • The --tools=pov option is important in order to generate PoVs. If you omit it, you will just get random test cases, which will at best crash the binary depending on what the constraint solver generates. Vulnerability generation involves a number of additional plugins that may add overhead to the analysis, so they are disabled by default unless you actually need them.

  • Finally, the @@ marker tells that the binary takes as input a file. This marker will be replaced at runtime with an actual file containing symbolic data.


You should build the desired guest image first, or use the -d flag if you want to download a pre-built image instead. Only Linux-based images are available for download. You must build Windows images yourself.

  1. Launch S2E and wait until the analysis completes (it should take less than 30 seconds).

$ cd ~/s2e/env/projects/vuln-lin32-32
$ ./

Look at the files inside ~/s2e/env/projects/vuln-lin32-32/s2e-last. This folder contains all the analysis results. You should have files that look like this:

$ ls ~/s2e/env/projects/vuln-lin32-32/s2e-last

These are concrete input files that demonstrate the vulnerability. Pick one of them and run it on the binary on your host machine under GDB, as follows.

  1. Confirm that the PoVs exercise the vulnerabilities

$ gdb --args ./vulnerabilities s2e-last/_tmp_input-recipe-type1_i386_generic_shellcode_eax.rcp@0x804883e-pov-type1-3
(gdb) r

The program should crash at program counter 0x44556677 and contain 0xccddeeff in the eax register.

Starting program: /home/user/s2e/env/projects/vuln-lin32-32/vulnerabilities

Demoing function pointer overwrite

Program received signal SIGSEGV, Segmentation fault.
0x44556677 in ?? ()
(gdb) info registers
eax            0xccddeeff       -857870593
ecx            0x44556677       1146447479
edx            0x40     64
ebx            0x0      0
esp            0xffffcd1c       0xffffcd1c
ebp            0xffffcd68       0xffffcd68
esi            0x804b008        134524936
edi            0xf7fa2000       -134602752
eip            0x44556677       0x44556677
eflags         0x10246  [ PF ZF IF RF ]
cs             0x23     35
ss             0x2b     43
ds             0x2b     43
es             0x2b     43
fs             0x0      0
gs             0x63     99

That’s it. What you have here is an input file that proves that an attacker has full control of the program counter as well as an additional general purpose register.

Understanding recipes

So where do the magic values 0x44556677 and 0xccddeeff come from? Short answer: from the FilePovGenerator plugin configuration. Open s2e-config.lua and look for the following section:

-- This plugin writes PoVs as input files. This is suitable for programs that
-- take their inputs from files (instead of stdin or other methods).
pluginsConfig.FilePovGenerator = {
    -- The generated PoV will set the program counter
    -- of the vulnerable program to this value
    target_pc = 0x0011223344556677,

    -- The generated PoV will set a general purpose register
    -- of the vulnerable program to this value.
    target_gp = 0x8899aabbccddeeff

You can specify in the configuration file which register values to set for the concrete test case. The configuration accepts 64-bit values, which the PoV generator will truncate if the program is 32-bit.

Now, in order to understand why these values get written into the program counter and the eax register, we need to look at recipes. If you look closer, you will notice that the test case file name above contains type1_i386_generic_shellcode_eax. This means that this test case was derived from the type1_i386_generic_shellcode_eax.rcp recipe. You can find the recipes in the recipes folder in your project directory. The recipe used here has the following contents:

# Set GP and EIP with shellcode
# mov eax, $gp
# mov ecx, $pc
# jmp ecx
[EIP+0] == 0xb8
[EIP+1] == $gp[0]
[EIP+2] == $gp[1]
[EIP+3] == $gp[2]
[EIP+4] == $gp[3]
[EIP+5] == 0xb9
[EIP+6] == $pc[0]
[EIP+7] == $pc[1]
[EIP+8] == $pc[2]
[EIP+9] == $pc[3]
[EIP+10] == 0xff
[EIP+11] == 0xe

Here is what every line of this recipe means:

  • reg_mask and pc_mask indicate which bits of the general purpose register and program counter can be controlled by the attacker. It will be 0xffffffff in almost all cases, meaning that all bits can be controlled.

  • type can be either 1 or 2 (see DARPA’s CGC terminology). Type 1 recipes control register values. Type 2 allow memory exfiltration.

  • arch and platform define when the recipe is applicable. In this case, we have a recipe that works on 32-bit programs and all OSes (Windows, Linux, etc…). If you have OS-specific shellcode, you must set the platform field accordingly. Please refer to the recipe plugin implementation to get an updated list of supported platforms.

  • gp means which general purpose register this recipe controls

  • exec_mem indicates which register must point to executable memory in order for the shell code to work. Executable memory is not required for more complex recipes that use ROP chains.

  • [EIP+0] == 0xb8 means that the first byte of the memory location referenced by EIP at the moment of exploitation must contain 0xb8. The value of EIP itself is not directly controlled by the recipe. Instead, the recipe plugin enumerates all suitable memory areas in the program’s address space and picks one area that satisfies the constraints (i.e., the area contains enough symbolic bytes that can be suitably constrained to generate the PoV). This also means that the chosen area must be fixed across program invocations. If the recipe plugin ends up choosing, say, an executable stack location, the resulting PoV may not be replayable (i.e., running it in GDB may not produce the desired crash). In practice, if the guest OS uses data execution prevention (DEP)s, you will need to encode a ROP chain in the recipe. The vulnerabilities demo allocates an executable area at a fixed location in order to simplify the recipe.

PoVs for DARPA Decree/CGC binaries

DARPA’s Cyber Grand Challenge (CGC) was the world’s first all-machine hacking tournament. S2E was a key component in CodeJitsu’s Cyber Reasoning System (CRS) and was used to automatically find vulnerabilities and exploit them. This demo walks you through the process of using S2E to find and generate a “proof of vulnerability” (PoV - i.e. an exploit) in a CGC challenge binary (CB).

The CGC Final Event (CFE) ran on the Decree operating system. Decree is a modified Linux OS with a reduced number of system calls. In addition to this, the Decree OS has been modified to add “hook points” for S2E (e.g. to signal process creation, termination, etc.) and to allow S2E to inject symbolic values. The source code for the Decree OS is available at

The current version of S2E does not use the old Decree kernel anymore. It was based on Linux 3.13, which does not build on modern Linux images that currently ship with S2E. Instead, S2E comes with cgcload, a tool that emulates the Decree system calls in user-mode using the syscall user dispatch mechanism in recent Linux kernels. This tool is a loader that maps the native CGC challenge binary into its address space, sets up syscall emulation that translates Decree syscalls into native Linux syscalls, then jumps to the binary’s entry point.

To run CGC binaries, build the 32-bit Linux image. This is the same image that is used for running normal Linux binaries.

$ cd ~/s2e/env
$ s2e image_build debian-12.5-i386

Next, create an analysis project for a challenge binary. Sample CBs are available here and can be built using the instructions here . The remainder of this tutorial will focus on the CADET_00001 program (a pre-compiled version of which is available here), but the ideas and techniques should be applicable to all of the CBs. Note that the S2E tooling does not support multi-binary CBs.

The following command creates a projects/CADET_00001 directory with various scripts and configuration files needed by S2E, as described here.

$ s2e new_project ./source/s2e/decree/samples/CADET_00001

Then run S2E:

$ cd projects/CADET_00001
$ ./

Wait for PoVs to be generated (s2e-last/*.c files) then interrupt S2E with Ctrl+C. Finally, verify that the generated PoVs are valid, i.e., they can reproduce the correct crashes when run, using the following command:

$ ./

Understanding CGC-style PoVs

If you followed the tutorial on PoV generation on Linux, you will notice that the PoV format for CGC binaries is different. Instead of being a concrete input file, S2E plugins produce PoVs in .c format. The reason for this is that CGC binaries read their input from stdin and write results to stdout. So in order to exercise the vulnerability, the PoV must implement a two-way communication with the program, by reading the program’s output and writing an appropriate input. This is different from file-based PoVs, where all the input is sent to the program at once, and the program’s output is ignored.


Many binaries, not just CGC binaries, use an interactive type of communication format, where input is read from stdin and results are written to stdout (e.g., command line utilities). S2E only supports file-based PoVs on Linux and Windows. Supporting interactive binaries for other platforms is work in progress.

For this reason, replaying a CGC-style PoV is more complex. It requires a special setup so that the PoV can communicate with the CB. The script takes care of all the steps. First, the PoV file must be compiled to a binary. Then the povtest tool takes this binary and connects its standard input/output to the challenge binary’s output/input. It negotiates the PoV parameters with the PoV binary and monitors the challenge binary for crashes. When the PoV terminates, povtest verifies that the challenge binary crashed at the correct location (type 1 PoV) or the PoV binary exfiltrated the correct data from the challenge binary’s address space (type 2 PoV).

# Build the PoV source code. The C file is generated by the DecreePoVGenerator plugin.
$ gcc -O0 -g -m32 -o s2e-last/recipe-type2_i386_decree_shellcode_1.rcp-pov-type2-1 \
      -L$S2EDIR/install/bin/guest-tools32/lib -I$S2EDIR/source/s2e/guest/linux/include \
      s2e-last/recipe-type2_i386_decree_shellcode_1.rcp-pov-type2-1.c -lpov -lcgc

# Verify the PoV.
$ $S2EDIR/install/bin/guest-tools32/povtest \
      --pov=s2e-last/recipe-type2_i386_decree_shellcode_1.rcp-pov-type2-1 \
      -- $S2EDIR/bin/guest-tools32/cgcload $S2EDIR/source/s2e/testsuite/pov-cgc-cadet0/CADET_00001/bin/CADET_00001


S2E does not use the original DARPA CGC tooling for PoV verification. The tooling is not maintained and does not run on modern Linux distributions, mostly because it uses an old Python 2 version. The Decree kernel does not build anymore either so it is not possible to run the CGC binaries natively.

Plugin architecture overview

In this section, we will give an overview of the plugins involved in generating a PoV. The diagram below summarizes the relationship between plugins.


S2E has an event-based plugin architecture. The execution engine exports a core set of low level events (declared in CorePlugin.h) to which plugins must subscribe if they want to do anything useful. The most important core events are related to guest instruction translation. Plugins must use them if they want to instrument guest code (e.g., to be notified when some instructions of interest are executed). Plugins may also define and export their own high-level events that other plugins can listen to. For example, an OS monitoring event could instrument the guest kernel so that it can notify other plugins about process and thread creation.

S2E plugins can be classified in roughly two sets: generic plugins and domain-specific plugins. Generic plugins can be seen as library functions in a programming language, which form building blocks for domain-specific plugins (i.e., the application/tool built on top of S2E). In the diagram above, the generic plugins take care of abstracting away low level details of guest execution (e.g., keeping track of processes running in the guest for a given execution state, building the memory and module map, etc). Domain-specific plugins rely on these generic plugins to simplify their implementation and focus on the problem to solve. For example, the recipe plugin can focus on monitoring vulnerable instructions in the processes of interest. It does not have to worry about figuring out to which process every instruction belongs. OS monitoring plugins take care of that.

S2E provides two ways to monitor guest execution: introspection or guest agents. Introspection consists of observing the execution stream or guest memory and react to OS-specific constructs without injecting code or modifying the guest in any way. For example, in order to get the current thread and process ID for Windows guests, the WindowsMonitor plugin probes several kernel data structures in memory. WindowsMonitor does this without involving the guest. A guest agent is a program that runs in the guest and uses the guest’s APIs in order to communicate relevant events to S2E plugins. For example, WindowsMonitor relies on the s2e.sys guest driver to monitor some OS events that would be difficult to catch with pure introspection.

In general, it is much simpler to use guest agents whenever possible. For example, in order to monitor Windows processes creation, it is simpler to write a guest driver that registers a callback through the PsSetCreateProcessNotifyRoutine kernel API and then calls S2E via s2e_invoke_plugin("WindowsMonitor", ...); instead of having WindowsMonitor figure out where Windows stores its data structures in memory and then trying to parse them. The difficulty is compounded by the fact that these structures are mostly undocumented and change with every OS revision.

Some tasks can be solved by combining introspection with guest agents. One example is getting the current process and thread ID from a plugin. Ideally, the plugin should be able to call the guest OS directly in order to get this information. However, the execution model of S2E only allows the guest calling plugins, not the reverse. So in order, to get these IDs, WindowsMonitor has to use introspection. However, it gets some help from the guest agent, which communicates to WindowsMonitor the locations of key Windows data structures in memory and offsets in these structures.

Plugins involved in PoV generation

In the previous section, we gave an overview of the S2E plugin architecture. Here we explain in more details the plugins that are involved in detecting vulnerabilities and generating PoVs.


This is the most important plugin for PoV generation. It monitors execution and looks for interesting symbolic addresses (assignment to program counters and memory reads/writes). When it finds one, it tries to constrain the address in such a way that leads to controlling registers and memory content according to the specification in the recipe. The recipe plugin supports Type 1 and Type 2 PoVs. PoV generation will not work without recipes.


Sometimes, the recipe plugin generates dozens of redundant PoVs that have only small variations. This plugin filters the PoVs generated by the recipe plugin to keep only those that are interesting. When the plugin decides that a PoV is worth keeping, it calls the actual PoV generation plugin, e.g., DecreePovGenerator or FilePovGenerator described below.


This plugin generates PoVs using the standard defined by DARPA. This PoV generator is designed for interactive programs that consist of a sequence of reads and writes to the standard input/output pair.


This plugin generates PoVs suitable for use by programs that take simple files as input and do not interact through the standard input/output.


This plugin collects PoVs and other interesting events and sends them to a backend server through a JSON API. This is useful if you want to integrate S2E into a cluster and monitor progress from a centralized console. We built this plugin for the CGC competition initially (hence the name), but it could be made more generic.

The plugins above are useful to generate PoVs. They do not help finding them. Finding vulnerabilities is the job of search heuristics, described below:


Seed files (or test inputs) are concrete inputs for the program under analysis. These files can be anything that the program accepts (PNG files, documents, etc.). They can be obtained from a fuzzer, generated by hand, etc. For CGC, seeds are binary executables compiled from XML of C PoV format.

The SeedSearcher plugin fetches seed files to concolically guide execution in the target program. Seed files are placed in the seeds directory. During analysis, the SeedSearcher plugins polls the seeds directory for new seeds. When it finds new seeds, the plugin forks a new state that fetches the new seed and then runs the binary using that seed as input.

Seed files can have different priorities. For example, if a fuzzer finds a seed that crashes the program, S2E may want to use that seed before others that, e.g., only cover new basic blocks. The priority of a seed is specified in its name. Seed files use the following naming convention:


The index specifies the order of the seed. SeedSearcher fetches seed files by increasing index number. Higher priorities are specified with higher integer. In a given batch of seeds, SeedSearcher will schedule those with the highest priority first.

When there are many seed files, it is advantageous to run S2E on multiple cores. In this mode, the SeedSearcher will automatically load balance available seeds across all available cores. For example, if there are 40 cores available, SeedSearcher will attempt to run 40 seeds in parallel.

The SeedSearcher plugin works in conjunction with the guest bootstrap file. The bootstrap file is built in such a way that state 0 runs in an infinite loop and forks a new state when a new seed is available. If there are no seed files, the bootstrap forks a state in which the program is run without seeds.

CUPA Searcher

This searcher implements the Class Uniform Path Analysis (CUPA) algorithm as described in this paper. It can work together with the SeedSearcher plugin.

The bootstrap script

The bootstrap script is a file called that the guest fetches from the host and executes. It contains instructions on how to execute the program under analysis. More detail can be found in the s2e-env documentation.

The CGC script slightly differs from Linux bootstraps. One key difference is that seeds will always be enabled for CGC projects, so the while loop in the execute function will exist even if you do not intend to use seed files. Note that it will not affect symbolic execution - the SeedSearcher (described above) will simply never schedule this state (state 0) for execution. This state will always exist, which means that even if S2E explores all execution paths in the CB (which is possible for simple binaries such as CADET_00001), the analysis will not end because not all states have been killed. The analysis must therefore be manually stopped.