function substitution attacks (sub attacks): altering function pointers between CFI-allowed targets.

SACK can automatically modify control data to construct sub attacks.

There are three main components:

• Oracle Constructor: Construct security-oriented oracles and produce pairs of program inputs and expected security behaviors.

• Target Collector: Dynamically record the indirect control-flow transfer (ICT) targets during benign execution.

• Attack Engine (built on AFL): Perform substitutions and security violation check, report successful sub attacks.

Please check our paper published on NDSS 2026 for more details.

We provide four pre-built Docker images for evaluating the three core components of our system: the Oracle Constructor, Target Collector, and Attack Engine.

Each image packages specific pre-built target programs and their corresponding oracles. This separation avoids interference caused by conflicting global settings in certain programs.

Each image shares a similar directory layout, with this repository located at /ae-sack inside the container.

• Oracle constructor

  • Can be tested in any of the provided images.
  • OPENAI_API_KEY and network connection required.

• Target collector and Attack engine

  • The test environment are packaged into separate Docker images.
  • Choose the image(s) that correspond to the target programs/oracles you want to evaluate.

This table outlines the relationship between each tested oracle, the location of its corresponding scripts, the Docker image used for evaluation, and the associated experiment folder.

• Step 1: Select Docker Image(s)

  • Choose one or more images based on the target components and oracles you wish to evaluate.

• Step 2: Pull the Docker Image

  docker pull sackae/sack_main:latest

• Step 3: Create and Start a Container

  docker run --cap-add=SYS_PTRACE \
             --name ae-sack-main \
             --security-opt seccomp=unconfined \
             -it sackae/sack_main:latest \
             tail -f /dev/null
  docker exec -it ae-sack-main /bin/bash           

• Step 4: Follow the Provided Instructions to Run the Experiments

In addition to pulling images from DockerHub, the Dockerfile, all Docker images, and this repository are archived on Zenodo at 🔗[https://doi.org/10.5281/zenodo.17782315] for permanent access.

The Zenodo archive includes the following OCI image tarballs. Here is the mapping to their corresponding Docker image tags:

To convert an OCI archive from Zenodo into a Docker-compatible archive, use:

apt-get install -y skopeo

# Convert oci-archive to docker-archive
skopeo copy oci-archive:sack_main_latest.oci.tar \
             docker-archive:sack_main_latest.docker.tar

# Load into docker
docker load -i sack_main_latest.docker.tar
# Example output:
#   Loaded image ID: sha256:<IMAGE_ID>

# Tag the image
docker tag sha256:<IMAGE_ID> sack_main:latest

# Then continue with Step 3 from the "How to Use" section.

• Experiment 1 (E1) - Claim (C1): Oracle Constructor
  Refer to the README.md in /ae-sack/oracle-generation/ for instructions.

• Experiment 2 (E2) - Claim (C2): Target Collector
  Refer to the README.md in /ae-sack/target-collector/ for detailed guidance on running target collection.

• Experiment 3 (E3) - Claim (C3): Attack Engine
  Follow the README.md in /ae-sack/scripts/ to execute and validate attack generation experiments.

• Zhechang Zhang zbz5352@psu.edu
• Hengkai Ye hengkai@psu.edu
• Song Liu songliu@udel.edu
• Hong Hu honghu@psu.edu

SACK: Systematic Generation of Function Substitution Attacks Against Control-Flow Integrity

@inproceedings{zhang:sack,
  title        = {{SACK: Systematic Generation of Function Substitution Attacks Against Control-Flow Integrity}},
  author       = {Zhechang Zhang and Hengkai Ye and Song Liu and Hong Hu},
  booktitle    = {Proceedings of the 33rd Annual Network and Distributed System Security Symposium (NDSS 2026)},
  month        = feb,
  year         = {2026},
  address      = {San Diego, CA},
}

We thank National Science Foundation (NSF) for supporting our work. This research is supported by NSF under grant CNS-2247652 and CNS-233984.