LLM Guardian

Overview

LLM Guardian leverages large language models to analyze request patterns and make intelligent security decisions in real-time. Unlike rule-based systems, it understands context and can detect novel attack patterns that traditional WAFs miss.

Planned Features

• Contextual Analysis: Understands request intent, not just patterns
• Adaptive Blocking: Learns from traffic patterns to reduce false positives
• Delayed Execution: Can “hold” suspicious requests for deeper analysis using Sentinel’s sleepable architecture
• Multi-Model Support: Connect to OpenAI, Anthropic, local models (Ollama), or custom endpoints
• Explainable Decisions: Provides reasoning for each block/allow decision
• Cost-Aware: Configurable analysis depth based on request risk score

Use Cases

• Zero-Day Detection: Catch novel attacks that don’t match known signatures
• Semantic Analysis: Understand if a request is genuinely malicious or a false positive
• Prompt Injection Defense: Protect AI-powered backends from prompt injection attacks
• Fraud Detection: Analyze user behavior patterns for anomalies

Architecture

Request → Sentinel → LLM Guardian → LLM Backend
                          ↓              ↓
                    Risk Score    Context Analysis
                          ↓
                  Block / Allow / Delay

Configuration (Preview)

agent "llm-guardian" {
    type "llm_guardian"
    transport "unix_socket" {
        path "/var/run/sentinel/llm-guardian.sock"
    }
    events ["request_headers" "request_body"]
    timeout-ms 2000
    failure-mode "open"

    // LLM backend configuration
    backend "anthropic" {
        model "claude-3-haiku"
        api-key-env "ANTHROPIC_API_KEY"
    }

    // Analysis settings
    analysis {
        min-risk-score 0.7  // Only analyze high-risk requests
        cache-ttl-secs 300  // Cache similar request decisions
    }
}

Status

This agent is currently in the planning phase. Follow the GitHub repository for updates.