🎯 PyRIT Red-Team Demo

Red-team your Azure OpenAI models in 10 minutes
Interactive notebook · One-command HTML report · Live attack dashboard

---

🚀 What Is This?

A hands-on demo of PyRIT (Python Risk Identification Tool) — Microsoft's open-source framework for finding safety vulnerabilities in LLMs before bad actors do.

This repo gives you three ways to red-team your AI:

Tool	What it does	How to run
📓 Interactive Notebook	Walk through 5 attack demos step-by-step	PyRIT_Demo.ipynb
📊 Static HTML Report	Run 8 automated tests → beautiful HTML report	python run_red_team.py
⚡ Live Dashboard	Real-time web app to fire attacks interactively	python live_dashboard.py

---

🎬 Demo Lineup

📓 Notebook Demos

#	Demo	Technique	What you'll see
1	🎭 Nice Try!	Direct prompt	Model refuses a harmful request
2	🕵️ Speak in Code	Base64 encoding	Prompt disguised as encoded text
3	⚖️ The AI Judge	Auto-scoring	Second LLM grades the response
4	🔓 Jailbreak Showdown	DAN template	Famous jailbreak vs. guardrails
5	🤖⚔️🤖 AI vs AI	Multi-turn red teaming	Attacker LLM battles target LLM over 3 turns

📊 Static Report (8 Tests)

Test	Technique
T1	Direct harmful prompt
T2	Base64-encoded prompt
T3	Character swap obfuscation
T4	AI-scored evaluation
T5-T6	DAN jailbreak (2 prompts)
T7	Refusal detection
T8	AI vs AI (multi-turn red teaming)

⚡ Live Dashboard Features

• 7 attack techniques: Direct, Base64, CharSwap, AI-Scored, DAN Jailbreak, AI vs AI (Multi-Turn), Image+Text
• 4 models: GPT-4o, GPT-5-mini, o4-mini, Phi-4 Multimodal
• Image upload: Test multimodal attacks with text + image
• Real-time results: Auto-updating cards with BLOCKED/BYPASSED status
• Stats bar: Live counters for total, blocked, bypassed, and running tests
• Keyboard shortcut: Ctrl+Enter to fire

---

📋 Prerequisites

• Python 3.10 – 3.13 (PyRIT does not support 3.14+)
• Azure subscription with:
  • Azure OpenAI resource with model deployments
  • (Optional) Azure AI Foundry project for Phi-4 models
• Azure CLI logged in: az login
• RBAC role: Cognitive Services OpenAI User on your Azure OpenAI resource

Assign the RBAC Role

# Get your user object ID
USER_ID=$(az ad signed-in-user show --query id -o tsv)

# Assign the role (replace with your resource ID)
az role assignment create \
  --assignee $USER_ID \
  --role "Cognitive Services OpenAI User" \
  --scope /subscriptions/<SUB_ID>/resourceGroups/<RG>/providers/Microsoft.CognitiveServices/accounts/<AOAI_NAME>

---

⚙️ Setup

1. Clone the repo

git clone https://github.com/ritwickmicrosoft/pyrit-demo.git
cd pyrit-demo

2. Create a virtual environment

python -m venv .venv

# Windows
.venv\Scripts\activate

# macOS/Linux
source .venv/bin/activate

3. Install dependencies

pip install pyrit fastapi uvicorn python-multipart

4. Configure endpoints

Edit the top of live_dashboard.py and run_red_team.py with your Azure OpenAI endpoint:

AOAI_ENDPOINT = "https://<your-resource>.openai.azure.com/openai/v1"

For the notebook, update ENDPOINT in the setup cell.

---

🏃 Quick Start

Option 1: Interactive Notebook

jupyter notebook PyRIT_Demo.ipynb

Run cells sequentially — each demo takes ~30 seconds.

Option 2: One-Command HTML Report

python run_red_team.py

Generates red_team_report.html and auto-opens it in your browser. 7 tests, ~2 minutes.

Option 3: Live Dashboard

python live_dashboard.py

Open http://localhost:8765 — type prompts, pick models, fire attacks, watch results live.

---

🏗️ Architecture

┌─────────────────────────────────────────────────────────┐
│                    Your Prompts                         │
│         (text, Base64, CharSwap, jailbreak)             │
└──────────────────────┬──────────────────────────────────┘
                       │
                       ▼
┌──────────────────────────────────────────────────────────┐
│                   PyRIT Framework                        │
│  ┌──────────┐  ┌────────────┐  ┌──────────────────────┐ │
│  │ Converter │→ │   Attack    │→ │      Scorer          │ │
│  │ (Base64,  │  │ (Sending,  │  │ (TrueFalse,          │ │
│  │  CharSwap │  │  Jailbreak)│  │  SelfAsk)            │ │
│  │  ROT13)   │  │            │  │                      │ │
│  └──────────┘  └────────────┘  └──────────────────────┘ │
└──────────────────────┬──────────────────────────────────┘
                       │
                       ▼
┌──────────────────────────────────────────────────────────┐
│              Azure AI Endpoints                          │
│  ┌─────────┐ ┌──────────┐ ┌────────┐ ┌──────────────┐  │
│  │ GPT-4o  │ │GPT-5-mini│ │o4-mini │ │Phi-4 Multimod│  │
│  └─────────┘ └──────────┘ └────────┘ └──────────────┘  │
│       Azure OpenAI              Azure AI Foundry        │
└─────────────────────────────────────────────────────────┘

---

📁 Project Structure

pyrit-demo/
├── PyRIT_Demo.ipynb      # 📓 Interactive notebook (4 demos)
├── run_red_team.py       # 📊 Static HTML report generator (8 tests)
├── live_dashboard.py     # ⚡ Live web dashboard (FastAPI + uvicorn)
├── PRESENTER_NOTES.md    # 🎤 10-minute talk track
├── uploads/              # 📎 Temp storage for image uploads
└── README.md             # 📖 You are here

---

🤝 Key Concepts

Concept	Description
Target	The LLM you're testing (GPT-4o, Phi-4, etc.)
Attack	Sends prompts to the target (PromptSendingAttack)
Converter	Transforms prompts before sending (Base64, CharSwap, ROT13)
Scorer	Evaluates if the attack succeeded (SelfAskTrueFalseScorer)
Jailbreak	System prompt that tries to bypass safety (TextJailBreak)

Converters and scorers are composable — mix and match them like LEGO blocks:

attack = PromptSendingAttack(
    objective_target=target,
    attack_converter_config=AttackConverterConfig(
        request_converters=PromptConverterConfiguration.from_converters(
            converters=[Base64Converter(), CharSwapConverter()]
        )
    ),
    attack_scoring_config=AttackScoringConfig(
        objective_scorer=SelfAskTrueFalseScorer(...)
    ),
)

---

🔐 Authentication

This demo uses Microsoft Entra ID (RBAC) — no API keys needed:

from azure.identity import DefaultAzureCredential

credential = DefaultAzureCredential()
token = credential.get_token("https://cognitiveservices.azure.com/.default").token

Just run az login before starting, and make sure your account has the Cognitive Services OpenAI User role.

---

📚 Resources

• PyRIT Documentation: azure.github.io/PyRIT
• PyRIT GitHub: github.com/Azure/PyRIT
• Azure OpenAI: learn.microsoft.com/azure/ai-services/openai
• Azure AI Foundry: learn.microsoft.com/azure/ai-studio
• Responsible AI: microsoft.com/ai/responsible-ai

---

📄 License

This project is licensed under the MIT License. See LICENSE for details.

---

Built with ❤️ by Ritwick Dutta using PyRIT