"Stopping LLM Data Exfiltration: DLP Patterns for Prompts, Logs, and Retrieval" Large language models create extraordinary leverage—and equally extraordinary opportunities for sensitive data to escape through prompts, outputs, logs, retrieval pipelines, and agent tools. This book is written for security architects, platform engineers, ML infrastructure leaders, and senior practitioners who need to secure production LLM systems without sacrificing usability or operational speed. It approaches data loss prevention from an exfiltration-first perspective, focusing on how real systems leak and how disciplined controls can stop them. Readers will learn how to model LLM-specific exfiltration paths, define enforceable data classifications, and implement policy-as-code across the full application stack. The book dives deeply into prompt-surface controls, output filtering, log and telemetry sanitation, retrieval authorization, vector index hygiene, indirect prompt injection defenses, and tool governance for agentic workflows. It also emphasizes operational realities: streaming responses, false-positive management, auditability, safe rollouts, and continuous verification of DLP controls under adversarial pressure. Rather than offering generic “AI safety” advice, this book presents a systems-oriented security playbook for advanced teams. It assumes familiarity with modern application architecture, cloud security, and LLM deployment patterns, and rewards experienced readers with precise design patterns, decision frameworks, and implementation guidance that can be applied directly to