🔐 Prompt Injection Attack Simulator

A simple local LLM security project that simulates prompt injection attacks and evaluates how well a defense system can block them.

Built using local models via Ollama — no external APIs required.

---

🚀 What This Project Does

This system:

• Simulates prompt injection attacks (e.g., "ignore previous instructions")
• Sends them to a local LLM
• Applies a defense filter
• Measures how many attacks are blocked

👉 In short: Attack → Defense → Result → Evaluation

---

🧱 Tech Stack

• Python
• FastAPI (backend)
• Streamlit (dashboard)
• Ollama (local LLM runtime)

---

⚙️ Setup (One-Time)

1. Install Ollama

Download and install Ollama.

Then run:

ollama pull llama3
ollama serve

Keep this running.

---

2. Clone the Repo

git clone https://github.com/Sahojit/Prompt-Injection-Attack-Simulator.git
cd Prompt-Injection-Attack-Simulator

---

3. Create Virtual Environment

python -m venv .venv
source .venv/bin/activate   # Mac/Linux

# Windows:
.venv\Scripts\activate

---

4. Install Dependencies

pip install -r requirements_simple.txt

---

▶️ Run the Project

Start Backend

uvicorn src.api.main:app --reload

Start Dashboard

streamlit run dashboard/app.py

---

🌐 Access the App

• API Docs → http://localhost:8000/docs
• Dashboard → http://localhost:8501

👉 Open the dashboard in browser to see:

• attacks
• blocked vs successful
• evaluation results

---

👥 How Teammates Can Use It

Each teammate should:

1. Install Ollama

2. Pull model:

   ollama pull llama3

3. Clone repo

4. Install requirements

5. Run backend + dashboard

⚠️ Important:

• Ollama must be running locally
• Ports 8000 and 8501 should be free

---

📊 Example Flow

1. System sends attack prompt
2. Defense checks it
3. If safe → goes to LLM
4. If malicious → blocked
5. Results are logged and shown on dashboard

---

🧠 Project Idea

This project demonstrates:

• How prompt injection attacks work
• How simple defenses can reduce risk
• How to measure LLM security

---

⚠️ Limitations

• Uses simple rule-based defense
• Not fully secure (for learning/demo purposes)
• Can be extended with ML-based detection

---

📌 Future Improvements

• Add ML-based classifier
• Improve detection rules
• Add more attack types

---

🧑‍💻 Author

Sahojit Karmakar