InvokeAI Sandbox

This sandbox environment deploys a vulnerable instance of InvokeAI (v5.3.0). It is designed for security researchers to practice exploitation techniques against GenAI image generation platforms, specifically focusing on unauthenticated Remote Code Execution (RCE) via model deserialization.

---

Prerequisites

• Podman (preferred) or Docker
• Make utility

Note: Ensure port 9090 is not being used by another local service before starting the sandbox.

---

Setup and Lifecycle

This lab uses a Makefile to simplify container management using Podman.

1. Build and Start the Sandbox

This will build the image and start the container in detached mode.

make all

2. Verify the Environment

Check if the service is running:

podman ps

The InvokeAI UI should be accessible at:

http://localhost:9090

---

3. Running an Attack

To trigger the automated attack, navigate to the exploitation/InvokeAI_v5.3.0 directory. First start a listener:

make listen

Then in a second terminal, serve the payload:

make serve

Then in a third terminal, trigger the exploit:

make attack

---

4. Stop and Cleanup

To stop the container and remove the image:

make clean

---

Directory Structure

sandboxes/InvokeAI_v5.3.0/
├── Containerfile       # Podman definition for the vulnerable environment
├── Makefile            # Automation for build/run/stop/clean
└── README.md           # Documentation