The Developer's Playbook for Large Language Model Security Bytom

Large language models (LLMs) are not just shaping the trajectory of AI, theyre also unveiling a new era of security challenges. This practical book takes you straight to the heart of these threats. Author Steve Wilson, chief product officer at Exabeam, focuses exclusively on LLMs, eschewing generalized AI security to delve into the unique characteristics and vulnerabilities inherent in these models. Complete with collective wisdom gained from the creation of the OWASP Top 10 for LLMs list-a feat accomplished by more than 400 industry experts-this guide delivers real-world guidance and practical strategies to help developers and security teams grapple with the realities of LLM applications. Whether youre architecting a new application or adding AI features to an existing one, this book is your go-to resource for mastering the security landscape of the next frontier in AI. Youll learn: Why LLMs present unique security challenges How to navigate the many risk conditions associated with using LLM technology The threat landscape pertaining to LLMs and the critical trust boundaries that must be maintained How to identify the top risks and vulnerabilities associated with LLMs Methods for deploying defenses to protect against attacks on top vulnerabilities Ways to actively manage critical trust boundaries on your systems to ensure secure execution and risk minimization Spis treści: Preface Who Should Read This Book Why I Wrote This Book Navigating This Book Section 1: Laying the Foundation (Chapters 13) Section 2: Risks, Vulnerabilities, and Remediations (Chapters 49) Section 3: Building a Security Process and Preparing for the Future (Chapters 1012) Conventions Used in This Book OReilly Online Learning How to Contact Us Acknowledgments 1. Chatbots Breaking Bad Lets Talk About Tay Tays Rapid Decline Why Did Tay Break Bad? Its a Hard Problem 2. The OWASP Top 10 for LLM Applications About OWASP The Top 10 for LLM Applications Project Project Execution Reception Keys to Success This Book and the Top 10 List 3. Architectures and Trust Boundaries AI, Neural Networks, and Large Language Models: Whats the Difference? The Transformer Revolution: Origins, Impact, and the LLM Connection Origins of the Transformer Transformer Architectures Impact on AI Types of LLM-Based Applications LLM Application Architecture Trust Boundaries The Model Public APIs: The convenience and the risks Privately hosted models: More control, different risks Risk considerations User Interaction Training Data Access to Live External Data Sources Access to Internal Services Conclusion 4. Prompt Injection Examples of Prompt Injection Attacks Forceful Suggestion Reverse Psychology Misdirection Universal and Automated Adversarial Prompting The Impacts of Prompt Injection Direct Versus Indirect Prompt Injection Direct Prompt Injection Indirect Prompt Injection Key Differences Mitigating Prompt Injection Rate Limiting Rule-Based Input Filtering Filtering with a Special-Purpose LLM Adding Prompt Structure Adversarial Training Pessimistic Trust Boundary Definition Conclusion 5. Can Your LLM Know Too Much? Real-World Examples Lee Luda GitHub Copilot and OpenAIs Codex Knowledge Acquisition Methods Model Training Foundation Model Training Security Considerations for Foundation Models Model Fine-Tuning Training Risks Retrieval-Augmented Generation Direct Web Access Scraping a specific URL Using a search engine followed by content scraping Example risks Accessing a Database Relational databases Vector databases Reducing database risk Learning from User Interaction Conclusion 6. Do Language Models Dream of Electric Sheep? Why Do LLMs Hallucinate? Types of Hallucinations Examples Imaginary Legal Precedents Airline Chatbot Lawsuit Unintentional Character Assassination Open Source Package Hallucinations Whos Responsible? Mitigation Best Practices Expanded Domain-Specific Knowledge Model fine-tuning for specialization RAG for enhanced domain expertise Chain of Thought Prompting for Increased Accuracy Feedback Loops: The Power of User Input in Mitigating Risks Clear Communication of Intended Use and Limitations User Education: Empowering Users Through Knowledge Conclusion 7. Trust No One Zero Trust Decoded Why Be So Paranoid? Implementing a Zero Trust Architecture for Your LLM Watch for Excessive Agency Excessive permissions Excessive autonomy Excessive functionality Securing Your Output Handling Common risks Handling toxicity Screening for PII Preventing unforeseen execution Building Your Output Filter Looking for PII with Regex Evaluating for Toxicity Linking Your Filters to Your LLM Sanitize for Safety Conclusion 8. Dont Lose Your Wallet DoS Attacks Volume-Based Attacks Protocol Attacks Application Layer Attacks An Epic DoS Attack: Dyn Model DoS Attacks Targeting LLMs Scarce Resource Attacks Context Window Exhaustion Unpredictable User Input DoW Attacks Model Cloning Mitigation Strategies Domain-Specific Guardrails Input Validation and Sanitization Robust Rate Limiting Resource Use Capping Monitoring and Alerts Financial Thresholds and Alerts Conclusion 9. Find the Weakest Link Supply Chain Basics Software Supply Chain Security The Equifax Breach Impact Lessons learned The SolarWinds Hack Impact Lessons learned The Log4Shell Vulnerability Impact Lessons learned Understanding the LLM Supply Chain Open Source Model Risk Training Data Poisoning Accidentally Unsafe Training Data Unsafe Plug-ins Creating Artifacts to Track Your Supply Chain Importance of SBOMs Model Cards Model Cards Versus SBOMs Purpose and focus Content Use in security and compliance Industry application CycloneDX: The SBOM Standard The Rise of the ML-BOM Building a Sample ML-BOM The Future of LLM Supply Chain Security Digital Signing and Watermarking Vulnerability Classifications and Databases MITRE CVE MITRE ATLAS Conclusion 10. Learning from Future History Reviewing the OWASP Top 10 for LLM Apps Case Studies Independence Day: A Celebrated Security Disaster Behind the scenes Chain of events Vulnerability disclosure 2001: A Space Odyssey of Security Flaws Behind the scenes Chain of events Vulnerability disclosure Conclusion 11. Trust the Process The Evolution of DevSecOps MLOps LLMOps Building Security into LLMOps Security in the LLM Development Process Securing Your CI/CD Implementing robust security practices Fostering a culture of security awareness LLM-Specific Security Testing Tools TextAttack Garak Responsible AI Toolbox Giskard LLM Scan Integrating security tools into DevOps Managing Your Supply Chain Protect Your App with Guardrails The Role of Guardrails in an LLM Security Strategy Input validation Output validation Open Source Versus Commercial Guardrail Solutions Mixing Custom and Packaged Guardrails Monitoring Your App Logging Every Prompt and Response Centralized Log and Event Management User and Entity Behavior Analytics Build Your AI Red Team Advantages of AI Red Teaming Red Teams Versus Pen Tests Tools and Approaches Red team automation tooling Red team as a service Continuous Improvement Establishing and Tuning Guardrails Managing Data Access and Quality Leveraging RLHF for Alignment and Security Conclusion 12. A Practical Framework for Responsible AI Security Power GPUs Cloud Open Source Multimodal Autonomous Agents Responsibility The RAISE Framework Limit your domain Balance your knowledge base Implement zero trust Manage your supply chain Build an AI red team Monitor continuously The RAISE Checklist Conclusion Index O autorze: Steve Wilson jest dyrektorem produktu w firmie Exabeam. Jest znaną postacią w społeczności sztucznej inteligencji i cyberbezpieczeństwa. Ma ponad 25-letnie doświadczenie w zakresie tworzenia platform oprogramowania dla największych firm technologicznych, takich jak Citrix Systems, Oracle i Sun Microsystems.