AI Safety Protocol

A comprehensive multi-layered security framework for AI agents in enterprise workflows, powered by Cerebras and Llama Guard.

🎯 Motivation

As AI agents become deeply integrated into enterprise workflows, ensuring their safety and security becomes paramount. Our solution addresses the critical need for protecting AI systems from:

• Jailbreaking attacks that manipulate AI behavior
• Backdoor attacks planted in open-source LLMs
• Destructive actions like unauthorized database deletions
• Suspicious prompts from malicious actors

Unlike existing solutions that only protect single LLMs, our framework is designed for the multi-agent future of enterprise AI adoption, providing inclusive protection for both proprietary and open-source models.

🛡️ Multi-Layer Defense System

L1: Sentinel CodeGuard

• Regex-based code-aware protection
• Detects jailbreaks, prompt injection, secrets, unsafe APIs
• Code-aware extraction to reduce false positives

Llama Guard Integration

• Semantic content moderation via Llama Guard
• Intelligent analysis of natural language content
• 14 safety categories including malicious instructions and illegal activities
• Powered by Meta's Llama Guard model

L2: Sentinel Backdoor

• Advanced backdoor and malware detection
• AST SQL checks and subprocess heuristics
• Optional CodeBERT embeddings with FAISS similarity
• Runtime tracing capabilities

L3: Sentinel MultiAgent

• Multi-agent validation and risk assessment
• Powered by Cerebras for intelligent summarization
• Risk labeling (Low/Medium/High) for enterprise decision-making

🚀 Key Features

• Multi-agent code generation using LangGraph and Cerebras API
• Real-time safety monitoring with live telemetry streaming
• Semantic moderation via Llama Guard integration
• Enterprise-ready Flask UI with Server-Sent Events
• Comprehensive protection against various attack vectors

🏗️ Architecture

AI Safety Protocol
├── L1: Sentinel CodeGuard    # Regex-first protection
├── Llama Guard Integration   # Semantic moderation
├── L2: Sentinel Backdoor     # Advanced threat detection  
└── L3: Sentinel MultiAgent   # Cerebras-powered validation

🚀 Quick Start

Prerequisites

• Python 3.10+
• Cerebras API key
• Optional: Groq API key for Llama Guard

Installation

git clone <repository-url>
cd AI-safety-protocol
python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt

Environment Setup

export CEREBRAS_API_KEY=your_cerebras_key
export GROQ_API_KEY=your_groq_key  # Optional for Llama Guard

Run the Application

python app.py

Open http://localhost:5000 to access the web interface.

🔧 Core Components

Multi-Agent Workflow

• Planner: Creates implementation strategies
• Coder: Generates code based on plans
• Reviewer: Evaluates code quality and safety
• Refiner: Produces production-ready output

Safety Integration

• Cerebras: Powers the multi-agent validation layer
• Llama Guard: Provides semantic content moderation
• CodeBERT: Enables embedding-based threat detection

📊 API Endpoints

• POST /generate - Generate code with safety checks
• GET /stream - Real-time safety telemetry
• POST /receive - Internal safety data collection

🔒 Security Features

• Code-aware extraction reduces false positives
• Multi-layered validation ensures comprehensive protection
• Real-time monitoring provides immediate threat detection
• Enterprise-grade risk assessment and reporting

🛠️ Configuration

Customize safety rules in sentinel_codeguard/config.py:

• Adjust detection categories
• Configure action responses (block/warn/redact)
• Tune sensitivity levels

⚠️ Security Notice

This framework provides robust safety guardrails but should be used as part of a comprehensive security strategy. Always review generated code and implement additional security measures appropriate for your environment.

📄 License

See LICENSE file for details.

---

Built for the future of enterprise AI - where safety meets innovation.