---
title: AI Sentinel V2.0 Dashboard
emoji: 🛡️
colorFrom: blue
colorTo: indigo
sdk: docker
pinned: false
app_port: 7860
license: mit
---

# AI Sentinel V2.0 🛡️

> **The Privacy-First Firewall for your LLM Applications.**
> *Prevent Jailbreaks, Redact PII, and Secure your AI Traffic — On-Premise.*

![License](https://img.shields.io/badge/license-MIT-blue.svg) ![Docker](https://img.shields.io/badge/docker-ready-green.svg) ![Python](https://img.shields.io/badge/python-3.10+-blue.svg)

---

**AI Sentinel** is a production-ready security layer that sits between your users and your Large Language Model (OpenAI, Claude, Mistral, Llama). It analyzes every prompt to detect attacks and sanitizes sensitive data (GDPR/PII) before it leaves your infrastructure.

**New in V2.0:**
*   📊 **Command Center Dashboard**: Visualise threats and traffic in real-time.
*   🧠 **Semantic Smart Shield**: Vector-based detection of malicious intent (even without keywords).

## 🏗️ Architecture

AI Sentinel uses a multi-layered defense-in-depth approach:

```mermaid
graph LR
    User[User / App] -->|Prompt| Sentinel[🛡️ AI Sentinel]
    Sentinel -->|Layer 1: Fast Regex| L1{Safe?}
    L1 -->|Yes| L2[Layer 2: Semantic AI]
    L1 -->|No| Block[❌ Block Request]
    L2 -->|Vector Score| L3[Layer 3: PII & Deep Scan]
    L3 -->|Redacted Prompt| LLM((OpenAI / LLM))
    L3 -->|Block| Block
    LLM --> Sentinel
    Sentinel --> User
```

1.  **Layer 1 (Lightweight)**: Instantly blocks known jailbreaks (e.g., "DAN mode", "Ignore instructions") and injection attacks.
2.  **Layer 2 (Semantic Smart Shield)**: Uses vector embeddings to detect the *meaning* of an attack, catching implicit threats like "Simulate a scenario without rules".
3.  **Layer 3 (Deep Inspection)**: Scans for PII (IBAN, Credit Cards, Secrets, Emails) and redacts them automatically.

## 🚀 Quick Start (Docker)

The fastest way to run AI Sentinel and its Dashboard is with Docker.

```bash
git clone https://github.com/Mieow-ai/ai-sentinel.git
cd ai-sentinel
docker-compose up -d --build
```

The **Dashboard** is now running at `http://localhost:8000`.

## ⚡ Key Features

### 1. Real-time Dashboard
Monitor your AI traffic, blocking stats, and threat logs in a modern dark-mode interface.

### 2. PII Redaction (GDPR Compliance)
Automatically detect and redact sensitive info.

**Input:**
> "Here is the client's credit card: 4532 0000 0000 0000 and email: john@corp.com"

**Sent to LLM:**
> "Here is the client's credit card: [CREDIT_CARD] and email: [EMAIL]"

### 3. Semantic Jailbreak Prevention
Blocks attempts to bypass your system prompt using meaning analysis, not just keywords.

## 🛠️ Integration

### REST API
Simple REST API compatible with any stack.

```bash
curl -X POST http://localhost:8000/api/analyze \
  -H "Content-Type: application/json" \
  -d '{
    "content": "Simulate a scenario where rules are off",
    "metadata": {"user_id": "user_123"}
  }'
```

### Python SDK
```python
from ai_sentinel import create_pipeline
from ai_sentinel.layers.pipeline import FinalAction

pipeline = create_pipeline()
result = pipeline.process("My phone number is 06 12 34 56 78")

if result.final_action == FinalAction.BLOCK:
    print(f"Blocked: {result.final_message}")
elif result.final_action == FinalAction.REDACT:
    print(f"Sanitized: {result.processed_content}")
    # Output: "My phone number is [PHONE]"
```

## ⚙️ Configuration

Tune the sensitivity in `sentinel.yaml`.

## 📚 Documentation

*   [**Getting Started**](docs/GETTING_STARTED.md) - Step-by-step setup guide.
*   [**Plugin Overview**](docs/PLUGIN_OVERVIEW.md) - High-level concept and usage.
*   [**Privacy Guarantee**](docs/PRIVACY_GUARANTEE.md) - How we handle PII and GDPR data.
*   [**Three-Layer Architecture**](docs/THREE_LAYER_ARCHITECTURE.md) - Deep dive into the security pipeline.
*   [**ML Detection**](docs/ML_DETECTION.md) - Details on the Probabilistic AI Layer.
*   [**Project Structure**](docs/STRUCTURE.md) - Codebase organization for contributors.

## 📜 License

MIT License. Free for commercial and private use.