AFL++ Seed Mutator

Fine-tuned Llama 3.1 8B model for generating AFL++ seed mutations. Trained on ancestry chains from fuzzing nginx.

Setup

1. Install Ollama

Download from ollama.com

2. Clone this repo

git clone https://github.com/VibeFuzzing/llm-mutator.git
cd llm-mutator

3. Download model parts

Download all 3 model parts from the Releases page:

• model_q4km.part0
• model_q4km.part1
• model_q4km.part2

Place them in the repo directory.

4. Reassemble the model

Windows (PowerShell):

.\reassemble_gguf.ps1 model_q4km.gguf

Linux/Mac:

chmod +x reassemble_gguf.sh
./reassemble_gguf.sh model_q4km.gguf

5. Create the Ollama model

ollama create afl-mutator -f Modelfile

Usage

ollama run afl-mutator

Provide an ancestry chain of seeds with metadata:

[id:0 depth:1 bitmap:406 favored:False new_cov:True]
GET / HTTP/1.1\r\n\r\n
---
[id:1 depth:2 bitmap:435 favored:True new_cov:True]
GET / HTTP/1.1\r
---
[id:67 depth:2 bitmap:432 favored:True new_cov:False]
GEP /\xfdGwT11/:\xb8TT:/12G/\x04

The model outputs a mutated seed that can be fed to AFL++.

How It Works

The model was trained on seed mutation lineages extracted from AFL++ fuzzing campaigns. Each training example contains an ancestry chain showing how seeds evolved through mutations, paired with a successful child seed that discovered new coverage or was favored by AFL++. The model learns to produce the next productive mutation given a history of prior seeds.