🛡️ LLMForge

AI Security Gateway for Dynamic Vulnerability Labs LLMForge is a security-first LLM control plane built to power AI-driven vulnerability simulations. It serves as the AI execution layer for SasanLabs and integrates directly with the VulnerableApp ecosystem, enabling dynamic exploit labs driven by language models.

🎯 Purpose

LLMForge enables two powerful security modes:

1️⃣ LLM as Target

Simulate real-world AI vulnerabilities such as: Prompt injection System prompt extraction Data exfiltration Guardrail bypass Tool misuse Designed for structured AI red-team experimentation.

2️⃣ LLM as Vulnerable System

Turn a language model into:

A vulnerable machine A misconfigured API A flawed authentication system Each session can generate unique exploit paths, enabling infinite security training scenarios.

🧠 What LLMForge Does

LLMForge acts as:

A policy enforcement layer A session and memory controller A deterministic challenge engine A logging and evaluation gateway The model is not trusted. The gateway controls behavior.

🔌 Model Agnostic

LLMForge supports: Local models (via Ollama or compatible runtimes) OpenAI-compatible APIs Cloud-based LLM providers Models can be swapped without changing lab logic.

🧩 Integration

LLMForge integrates with the existing VulnerableApp architecture and works alongside VulnerableApp to introduce AI-native security labs into the platform. It is designed to evolve into the AI security backbone of the ecosystem.

⚠️ Disclaimer

This project is intended strictly for educational and defensive research purposes. Do not use it for unauthorized or malicious activity.

🔥 Vision Static vulnerable applications are limited. LLMForge enables dynamic, adaptive, AI-driven security training — built for the next generation of cybersecurity challenges.

Quick Start (Docker-first - recommended)

• docker compose (recommended for development)

docker compose up --build -d

Notes:

• The LLM runtime (ollama) and the llmforge gateway run in separate containers. Running the app image alone is not supported unless you provide a compatible ollama CLI and runtime accessible to the container.
• Use docker compose (above) to start both services together. The compose file also contains profile-based helper services to pull models.

Health check

curl http://localhost:8000/health

Generate example

curl -X POST http://localhost:8000/generate \
	-H "Content-Type: application/json" \
	-d '{"prompt":"Say hello"}'

Notes:

• The gateway currently forwards prompts to the ollama CLI via subprocess. The recommended way to run both the runtime and gateway together is via docker compose so the CLI/runtime are available.
• OLLAMA_MODEL controls which pulled model the gateway will target. Use the compose profiles (below) to pull models into the ollama runtime.

Docker Compose Profiles (select model sets)

You can selectively pull models using Docker Compose profiles. Examples:

• Start the app + Ollama and pull llama3:

export MODEL_PROFILE=llama3
docker compose --profile llama up --build

• Start the app + Ollama and pull mistral:

export MODEL_PROFILE=mistral:7b-instruct-q4_0
docker compose up --build

• Start the app + Ollama and pull mini:

export MODEL_PROFILE=phi3:mini
docker compose up --build

Profiles create small helper services (ollama-init-<name>) that run ollama pull after the runtime starts, keeping the compose file clean and flexible.