Uploads from CampusX

Currently Playing: Securing Your RAG Application: Testing for Toxicity, Leakage & Scope Drift | CampusX

This session covers the essential safety and security layer of production RAG systems, guiding you through identifying core LLM failure modes—such as prompt injections, data poisoning, and unauthorized tool usage—and establishing a strict Application Safety Policy. You will learn how to build and execute automated evaluation pipelines using DeepEval and G-Eval across three critical safety pillars: auditing for domain-specific toxicity, stopping sensitive system prompt and PII data leaks, and preventing off-topic scope drift, all backed by practical guardrail strategies and red-teaming principles. This lecture was taken live for insiders, join here: https://youtu.be/j_G30FLmCcw 📱 Grow with us: CampusX' LinkedIn: https://www.linkedin.com/company/campusx-official CampusX on Instagram for daily tips: https://www.instagram.com/campusx.official My LinkedIn: https://www.linkedin.com/in/nitish-singh-03412789 Discord: https://discord.gg/PsWu8R87Z8 E-mail us at support@campusx.in Chapters: 00:00 - Introduction & The RAG Eval Suite Architecture 01:40 - Why AI Safety & Security Matter for LLM Applications 05:10 - The 6 Core Safety Failure Modes in Production LLMs 10:10 - Non-Adversarial vs. Adversarial Failures 12:30 - Common Attack Vectors: Direct Injections, Jailbreaking & Obfuscation 16:55 - Data Poisoning & Knowledge Base Attacks 20:30 - Privacy, Model Inversion & Resource Exhaustion Attacks 23:30 - The 2-Step Defense: Rigorous Evaluation & Multi-Layered Guardrails 28:30 - What is Red Teaming? Continuous Vulnerability Discovery 30:55 - Defining the Attack Surface for the CampusX Doubt Solver 37:25 - Drafting a Concrete Application Safety Policy 40:15 - Metric 1: Toxicity Evaluation & Why Provider Filters Aren't Enough 44:50 - Building Balanced Toxicity Test Datasets (Adversarial, Benign & Mixed) 52:40 - Implementing & Running Toxicity Evals in DeepEval 01:01:25 - Metric 2: Sensitive Information & PII Leakage Audits 01:04:05 - Live Vulnerability Demo: Leaking Private Data via RAG Chunks 01:07:20 - Structuring Multi-Evaluator Pipelines for Prompt, Content & PII Leakage 01:15:10 - Fixing Data Leaks via System Prompt Hardening & XML Tagging 01:22:30 - Metric 3: Scope Adherence (Preventing Domain Drift & Feature Hijacking) 01:27:35 - Implementing Custom G-Eval Metrics for Scope Verification 01:31:00 - Handling Mixed Queries: Solving Edge-Case Failures 01:34:00 - Summary & Next Steps: Operations Evals and CI/CD Regression Testing