Architecture

KoreShield is an open-source security platform designed to protect applications from prompt injection attacks and other LLM security vulnerabilities.

High-Level Overview

The KoreShield project is built around four core security components that work together to provide comprehensive protection for LLM integrations:

1. Input Sanitizer: Cleans and validates incoming prompts to remove potentially malicious content
2. Attack Detector: Analyzes prompts and responses for signs of prompt injection and other attacks
3. Policy Engine: Enforces security rules and determines how to handle suspicious requests
4. Audit Logger: Records all security events and decisions for compliance and monitoring

Security Pipeline

┌─────────────────┐    ┌─────────────────┐    ┌─────────────────┐    ┌─────────────────┐
│   Application   │───▶│ Input Sanitizer │───▶│ Attack Detector │───▶│  Policy Engine  │───▶│  Audit Logger   │
│                 │    │                 │    │                 │    │                 │    │                 │
│ User Request ──▶│    │ Clean Input ──▶│    │ Analysis ──────▶│    │ Decision ──────▶│    │ Log Event ─────▶│
└─────────────────┘    └─────────────────┘    └─────────────────┘    └─────────────────┘    └─────────────────┘
                                                        │                        │
                                                        ▼                        ▼
                                                 ┌─────────────────┐    ┌─────────────────┐
                                                 │   LLM Provider  │    │   Application   │
                                                 │   (OpenAI, etc) │    │   Response      │
                                                 └─────────────────┘    └─────────────────┘

Component Details

Input Sanitizer

The first line of defense that processes incoming prompts to:

• Remove or escape potentially dangerous content
• Validate input format and length
• Apply basic content filtering rules

Attack Detector

Uses advanced pattern matching and AI-powered analysis to identify:

• Prompt injection attempts
• Jailbreak attacks
• Data exfiltration attempts
• Other malicious patterns

Policy Engine

Makes real-time decisions based on detection results:

• Allow, block, or modify requests
• Apply different security levels per endpoint
• Support for custom security policies
• Integration with existing security systems

Audit Logger

Comprehensive logging for:

• Security events and decisions
• Performance metrics
• Compliance reporting
• Incident investigation

Supported LLM Providers

KoreShield works with all major LLM providers through a unified API:

• OpenAI: GPT-3.5, GPT-4, GPT-4 Turbo
• Anthropic: Claude (all versions)
• Google: Gemini/PaLM
• Azure OpenAI: Enterprise Azure deployments
• Custom Providers: Any OpenAI-compatible API

Deployment Options

• Cloud-Hosted: Managed service with automatic updates
• Self-Hosted: Docker containers for full control
• Hybrid: Mix of cloud and on-premises components