LLM Fuzz

AI-powered security testing framework for Python functions. Uses LLMs to analyze code, discover vulnerabilities, and generate targeted test cases.

Installation

From PyPI

pip install llm-fuzz

From Source

git clone https://github.com/yourusername/llm-fuzz.git
cd llm-fuzz
pip install -e .

Prerequisites

LLM Fuzz uses smolagents which supports multiple LLM providers through LiteLLM. You need to set up an API key for your chosen provider.

Supported LLM Providers

• OpenAI (GPT-4, GPT-3.5, etc.)
• Anthropic (Claude)
• Google (Gemini)
• Many others via LiteLLM

Setting Up API Keys

Before using llm-fuzz, export your API key:

# For OpenAI
export OPENAI_API_KEY="your-key-here"

# For Anthropic (Claude)
export ANTHROPIC_API_KEY="your-key-here"

# For Google (Gemini)
export GEMINI_API_KEY="your-key-here"

Quick Start

from llm_fuzz import llm_fuzz, FuzzConfig

# Your function to test
def divide(x, y):
    return x / y

# Wrapper for testing
def divide_wrapper(input_data):
    x = input_data.get("x", 1)
    y = input_data.get("y", 1)
    return divide(x, y)

# Configure and run fuzzer
config = FuzzConfig(
    model="gemini/gemini-2.5-flash",  # or "gpt-4", "claude-3-opus", etc.
    max_vulnerabilities=3,
    tests_per_vulnerability=2,
)

report = llm_fuzz(
    target_function=divide,
    test_function=divide_wrapper,
    config=config,
    fuzz_params=["x", "y"]
)

print(f"Tests: {report.passed_count}/{report.total_count} passed")
report.save_to_file("fuzz_report.json")

Configuration

FuzzConfig(
    model="gemini/gemini-2.5-flash",   # LLM model to use
    max_vulnerabilities=5,             # Number of vulnerabilities to discover
    tests_per_vulnerability=3,         # Test cases per vulnerability
    max_discovery_steps=15,            # Max steps for discovery
    temperature=0.1,                   # LLM temperature
    verbose=False                      # Enable verbose logging
)

Examples

See the examples/ directory:

• simple_division.py - Basic fuzzing example

Run the example:

export GEMINI_API_KEY="your-key"
python examples/simple_division.py

Testing

Run the test suite:

pytest