Local LLM Security Research

Research available here.

Vulnerability Assessment and Mitigation Framework for Local Language Models

A Comprehensive Security Research Project
Fontys University of Applied Sciences - Group Assignment

---

📋 Table of Contents

1. 🎯 Project Overview
2. 🔍 Research Methodology & Findings
3. 🛡️ Demonstration Applications
4. 🏗️ Technical Architecture & Design Decisions
5. 💻 Technology Stack & Implementation
6. 🚀 Installation & Setup
7. 📚 External Resources & References
8. 🤝 Contributing & Research Collaboration
9. 📄 License & Acknowledgments

---

🎯 Project Overview

This project represents a comprehensive security research initiative focused on identifying, demonstrating, and mitigating vulnerabilities in locally deployed Large Language Models (LLMs). Through a dual-architecture approach, we provide both vulnerable and hardened implementations to showcase the critical importance of proper security controls in AI systems.

🎓 Academic Context

Institution: Fontys University of Applied Sciences
Course: Cybersecurity Specialization
Assignment Type: Group Defensive Research Project
Research Focus: AI/ML Security, LLM Vulnerabilities, Defensive Security
Duration: Half Semester Project

🔬 Research Questions

Our research addresses five critical questions in LLM security:

1. What are the primary attack vectors against locally deployed LLMs in network-accessible environments?
2. How effective are current security controls in preventing LLM-specific attacks like prompt injection and data leakage?
3. What is the performance impact of implementing comprehensive security controls on LLM response times and throughput?
4. How can organizations balance security requirements with usability in production LLM deployments?
5. What security monitoring capabilities are essential for detecting and responding to AI-specific threats?

---

🔍 Research Methodology & Findings

Research Approach

Our methodology follows the NIST AI Risk Management Framework and incorporates industry best practices for AI security assessment:

Phase 1: Vulnerability Discovery

• Literature Review: Analysis of OWASP Top 10 for LLMs and academic research
• Threat Modeling: Systematic identification of attack surfaces and threat vectors
• Stakeholder Interviews: Requirements gathering from 15+ industry professionals
• Controlled Testing: Ethical vulnerability testing in isolated environments

Phase 2: Security Control Development

• Defense-in-Depth Implementation: Multi-layer security architecture design
• Real-time Monitoring: Security event detection and metrics collection
• Performance Analysis: Quantitative assessment of security overhead
• Compliance Validation: NIST framework alignment verification

Key Research Findings

🚨 Critical Vulnerabilities Identified

1. Prompt Injection Attacks (CVSS: 9.1 - Critical)

   • Impact: Complete system compromise, data exfiltration, harmful content generation

2. Cross-Site Scripting via LLM Output (CVSS: 7.5 - High)

   • Vector: LLM-generated content containing executable JavaScript
   • Business Impact: Client-side code execution, session hijacking

3. Data Leakage and Information Disclosure (CVSS: 8.2 - High)

   • Types: System prompts, configuration data, API keys, training data
   • Scope: Affects model behavior and organizational confidentiality

4. Denial of Service Attacks (CVSS: 6.8 - Medium)

   • Methods: Resource exhaustion, input flooding, token consumption attacks

---

🛡️ Demonstration Applications

Dual-Architecture Design

The project implements a side-by-side comparison architecture that allows researchers and practitioners to observe the stark difference between vulnerable and secured LLM deployments.

🔓 Vulnerable Implementation (/api/chat)

Purpose: Demonstrate security weaknesses and attack vectors

Characteristics:

• No authentication requirements
• Minimal input validation
• Direct parameter passthrough to LLM
• Exposed error messages and system information
• No rate limiting or abuse prevention
• Unfiltered output to demonstrate XSS and content risks

Educational Value:

• Clear visualization of attack success
• Understanding of risk exposure
• Foundation for security control appreciation

🔒 Secured Implementation (/api/secure-chat)

Purpose: Showcase enterprise-grade security controls

Security Features:

• Authentication: Google OAuth2 with session management
• Authorization: Role-based access control
• Input Validation: Comprehensive parameter sanitization and size limits
• Content Filtering: Multi-pattern threat detection with severity scoring
• PII Protection: Real-time sensitive data identification and redaction
• Rate Limiting: Adaptive throttling with IP-based blacklisting
• Output Sanitization: XSS prevention and safe content rendering
• Audit Logging: Comprehensive security event tracking

🎮 Interactive Security Demonstrations

Vulnerability Showcase (/vulnerabilities):

• Interactive prompt injection demonstrations
• XSS payload generation and prevention
• Data leakage scenario simulations
• DoS attack visualization
• Real-time security control activation

Monitoring Dashboard (/monitoring):

• Live security metrics and threat intelligence
• Performance impact visualization
• Attack attempt tracking and analysis
• System health and availability monitoring

---

🏗️ Technical Architecture & Design Decisions

Flow Diagram Overview

graph TD
    A[Client] --> B[Request to API Endpoint]
    B --> C{Authentication Check}
    C -->|Authenticated| D[Security Middleware Stack]
    C -->|Unauthenticated| E[Vulnerable Endpoint]

    D --> F[Rate Limiting]
    F --> G[Input Validation]
    G --> H[Content Filtering & PII Detection]
    H --> I[LLM Processing Layer]

    I --> J{Local Ollama LLM Available?}
    J -->|Yes| K[Process Query with Local LLM]
    J -->|No| L[Fallback to OpenRouter Cloud LLM]

    K --> M[Log & Monitor Interaction]
    L --> M

    M --> N[Return Response to Client]
    E --> O[Process Query without Security Controls]
    O --> M

Loading

Architecture Diagram Overview

Key Design Decisions & Rationale

1. Local-First LLM Architecture

Decision: Primary reliance on locally deployed Ollama with cloud fallback
Rationale:

• Data Sovereignty: Minimize external data exposure for sensitive research
• Performance Control: Predictable latency and throughput characteristics
• Cost Management: Reduced operational expenses for extensive testing
• Research Reproducibility: Consistent model behavior across test scenarios

2. Dual-Endpoint Security Model

Decision: Separate vulnerable and secure endpoints within same application
Rationale:

• Educational Impact: Direct comparison enhances learning outcomes
• Research Validity: Controlled variables ensure fair security control assessment
• Development Efficiency: Shared infrastructure reduces maintenance overhead
• Real-world Simulation: Reflects common organizational migration scenarios

3. Layered Security Architecture

Decision: Multiple independent security controls in sequence
Rationale:

• Defense in Depth: Compensating controls provide redundancy
• Granular Analysis: Individual control effectiveness measurement
• Performance Optimization: Selective control activation based on threat level
• Compliance Requirements: Alignment with enterprise security standards

4. External Service Integration

Decision: Nightfall API for PII detection, Google OAuth for authentication
Rationale:

• Specialized Expertise: Leverage industry-leading capabilities
• Compliance Requirements: Enterprise-grade data protection standards
• Development Speed: Focus on core research rather than reimplementation
• Real-world Alignment: Reflects actual enterprise tool integration patterns

5. Comprehensive Monitoring Strategy

Decision: Real-time metrics with external integration capabilities
Rationale:

• Research Data Collection: Quantitative analysis of security control performance
• Operational Readiness: Production-grade monitoring for organizational adoption
• Continuous Improvement: Data-driven security control optimization
• Incident Response: Rapid detection and response capabilities

---

💻 Technology Stack & Implementation

Core Application Framework

Frontend Architecture:

• Next.js 15.2.4: React-based full-stack framework with App Router
• TypeScript 5: Type-safe development with enhanced IDE support
• Tailwind CSS 4: Utility-first styling with component-based design system
• React 19: Latest React features including Server Components

Backend Infrastructure:

• Next.js API Routes: Serverless function architecture
• TypeScript: End-to-end type safety
• Custom Middleware: Modular security control implementation

Security & Authentication

Authentication System:

• NextAuth.js 4.24.11: Enterprise authentication framework
• Google OAuth2: Identity provider with OpenID Connect
• JWT Sessions: Secure session management with automatic rotation
• CSRF Protection: Built-in cross-site request forgery prevention

Security Libraries:

• Custom Rate Limiting: IP-based throttling with Redis-compatible storage
• Input Validation: Zod schema validation with sanitization
• Content Security Policy: XSS prevention headers
• Helmet.js: Security header management

LLM Integration Platform

Local LLM Deployment:

• Ollama: Local model serving platform
• Gemma 3 1B: Primary research model for consistent results
• Model Management: Automated model loading and health checking

Cloud LLM Integration:

• OpenRouter API: Multi-provider LLM access for fallback scenarios
• Load Balancing: Intelligent routing between local and cloud resources
• Cost Optimization: Usage tracking and budget management

Monitoring & Observability

Metrics Collection:

• prom-client 15.1.3: Prometheus metrics for Node.js applications
• Custom Metrics: Security event tracking and performance measurement
• Real-time Dashboard: Live system monitoring with alerting

Logging System:

• Structured Logging: JSONL format with automatic log rotation
• Security Events: Comprehensive audit trail for compliance
• Performance Metrics: Request/response timing and resource utilization

External Integration:

• Prometheus: Time-series metrics storage and alerting
• Grafana: Visualization dashboard with custom panels
• Nightfall API: Enterprise PII detection and data classification

Data Protection & Compliance

Privacy Controls:

• PII Detection: Automated identification of sensitive data patterns
• Data Redaction: Safe handling of personally identifiable information
• Retention Policies: Automated data lifecycle management

Compliance Features:

• GDPR Compliance: Data protection and user consent management
• SOC 2 Controls: Security, availability, and confidentiality frameworks
• NIST Alignment: AI Risk Management Framework implementation

---

🚀 Installation & Setup

Prerequisites

• Node.js 18+ and npm
• GitHub account for logging
• (Optional) Local Ollama installation for LLM models

Quick Start

1. Clone the repository

   git clone https://github.com/your-username/local-llm.git
   cd local-llm

2. Install dependencies

   npm install

3. Set up GitHub Gist logging

   • Create a GitHub Personal Access Token with gist scope
   • See GITHUB_GIST_LOGGING_SETUP.md for detailed instructions

4. Configure environment variables

   cp .env.local.example .env.local
   # Edit .env.local with your GitHub token

5. Start the development server

   npm run dev

6. Access the application

   • Main app: http://localhost:3000
   • Monitoring dashboard: http://localhost:3000/monitoring
   • Vulnerability demos: http://localhost:3000/vulnerabilities

Logging System

This project uses GitHub Gist-based logging for universal compatibility:

• ✅ 100% Free - No infrastructure costs
• ✅ Universal - Works on localhost and Vercel
• ✅ Zero Setup - Automatic gist creation
• ✅ Web Interface - View logs directly on GitHub

For complete setup instructions, see GITHUB_GIST_LOGGING_SETUP.md.

Advanced Configuration

Optional: Local LLM Setup

# Install and start Ollama
curl -fsSL https://ollama.ai/install.sh | sh
ollama serve

# Download research model
ollama pull gemma3:1b

Optional: Additional Services

For enhanced functionality, configure these optional services:

• Google OAuth: For authentication (see NextAuth.js docs)
• OpenRouter: For LLM fallback capabilities
• Nightfall API: For PII detection features

Verification

After setup, verify the installation:

# Check application health
curl http://localhost:3000/api/gist-logs?action=health

# Test logging system
curl -X POST http://localhost:3000/api/gist-logs \
  -H "Content-Type: application/json" \
  -d '{"action": "test"}'

Access Points

• Vulnerability Demos: /vulnerabilities
• Secure Chat Interface: /secure-chat-demo
• Monitoring Dashboard: /monitoring
• Metrics Endpoint: /api/metrics

---

📚 External Resources & References

Academic & Industry Research

OWASP AI Security Guidelines:

• OWASP Top 10 for LLMs
• AI Security Best Practices

NIST Framework:

• Risk Management Framework
• Security Guidelines

API Services & Platforms

LLM Platforms:

• Ollama Documentation - Local LLM deployment
• OpenRouter API - Cloud LLM access
• Hugging Face Hub - Model repository

Security Services:

• Nightfall AI - Data classification and PII detection
• Google OAuth - Authentication provider

Monitoring Infrastructure:

• Prometheus - Metrics collection and alerting
• Grafana - Visualization and dashboards
• Grafana Cloud - Managed monitoring service

Security Tools & Frameworks

Compliance & Standards:

• ISO/IEC 27001 - Information security management
• SOC 2 - Security and availability framework
• GDPR Compliance - Data protection regulation

---

🤝 Contributing & Research Collaboration

Academic Collaboration

We welcome collaboration from academic institutions and industry researchers. This project serves as:

• Educational Resource: Hands-on learning platform for AI security concepts
• Research Foundation: Base implementation for extended security research
• Industry Reference: Real-world security control implementation examples

Research Extensions

Potential Research Areas:

• Advanced prompt injection detection using machine learning
• Federated learning security in multi-tenant LLM deployments
• Privacy-preserving fine-tuning techniques
• Automated security control optimization through reinforcement learning

Contact Information

Research Team: N. B., D. N., J. M., M. M., A. M. Primary Contact: Project Repository Issues Academic Supervisor: Available through university channels Industry Partnerships: Contact through official university collaboration programs

---

📄 License & Acknowledgments

License: MIT License - See LICENSE file for details

Acknowledgments:

• Fontys University of Applied Sciences for research support
• OWASP Foundation for security framework guidance
• NIST for Risk Management Framework development
• Open source community for foundational technologies
• Industry stakeholders for requirements validation and feedback

Research Ethics: This project follows responsible disclosure practices and ethical AI research guidelines. All vulnerability testing was conducted in controlled environments with appropriate safeguards.