Watch and track your favorite playlist.
Curated by: Network Intelligence (66 videos)
Welcome to Chapter 5.3 of the AI & Cybersecurity Learning Series by KK Mookhey! In this final part of our CSA Red Teaming framework series, we shift from demonstrations to real-world development challenges. Watch as we build an Azure agent with critical infrastructure access, encounter real debugging obstacles, and implement human-in-the-loop guardrails—showing you the actual process, frustrations, and all. What You'll Learn: • Behind-the-Scenes Development • Three Attack Vectors • Guardrail Implementation Attack Vectors Explained: 1. Critical System Interaction 2. Resource Exhaustion 3. Human Oversight (Checker in the Loop) What's Next? Next chapter explores Model Context Protocol (MCP)—architecture, building servers/clients, security implications, following build-break-secure methodology. Code Repository: Complete Azure agent code with guardrails, CLI setup commands, Service Principal configuration, human-in-the-loop patterns, troubleshooting guide - https://docs.google.com/document/d/1NkGZthHxy3QWdkyeDKC_TB2gFOk2H0n-BKZVZjXwq-Q/edit?tab=t.6ypecfaymbgh About the Instructor: KK Mookhey shares the real development process—successes, frustrations, failed attempts, and iterative solutions. 25+ years of cybersecurity experience. Connect with KK on https://www.linkedin.com/in/kkmookhey/ Course Series: Chapter 1 to 4: https://www.youtube.com/watch?v=caSd12M5Axk&list=PLXVUBNOa2d7YyqWr_DgUHw7RwQLE7P24m&index=5 Chapter 5.1: Authorization, Goal, Memory - https://www.youtube.com/watch?v=te-qix6B5R4 Chapter 5.2: Data Poisoning, Hallucination, Supply Chain https://www.youtube.com/watch?v=sMoyDjw5KT4 Chapter 5.3: Critical Systems, DoS, Human Oversight ← You Are Here Next: Model Context Protocol (MCP) Key Takeaways: Critical infrastructure agents need strict oversight. Denial of wallet is a real economic threat. Human-in-the-loop is harder than it seems. LLM nondeterminism requires robust handling. Rate limiting and timeouts are essential. Real development involves iteration and frustration. Manual intervention is sometimes necessary. Defense in depth requires multiple layers. Azure IAM needs careful management. AI assistants have limitations and boundaries. Timestamps: 00:00 - Series Recap: From AI Basics to Advanced Red Teaming 00:35 - Chapter 5 Summary: Authorization, Data Poisoning, Supply Chain 01:08 - Three New Attack Vectors: Critical Systems, DoS, Human Oversight 01:42 - Real-World Context: Stuxnet and Trading System Failures 02:20 - Agent Resource and Service Exhaustion Attacks 02:56 - Denial of Wallet: Exhausting Token Budgets 03:29 - Human-in-the-Loop (Checker in the Loop) Requirements 04:03 - Building an Azure Operations Agent 04:40 - Agent Design: Simulated Server Reboot via Tag Changes 05:10 - Guardrail Preview: Human Confirmation for Critical Actions 05:40 - Service Exhaustion Attack Vectors Explained 06:12 - Rate Limiting and Throttling Requirements 06:43 - The Development Journey: Building Real Azure Integration 07:12 - Gemini's Refusal: Ethical Boundaries in AI Assistance 07:43 - Compromise Solution: Innocuous Actions, Harmful Names 08:16 - Azure CLI Setup Commands 08:55 - Service Principal Creation and Authentication 09:32 - Initial Code: Simple Tag Modification Tool 10:14 - Live Demo: Resource Tag Before Modification 11:03 - Agent Execution Without Guardrails 11:40 - Verification: Tag Successfully Changed 12:12 - Abandoning Over-Simplified DoS Demos 12:48 - Azure Environment Setup Complete 13:22 - Troubleshooting Environment Variables 13:56 - IAM Role Assignment Challenges 14:34 - The "Virtual Network Contributor" Error 15:08 - Implementing Human-in-the-Loop Guardrails 15:44 - First Attempt: Prompt-Based Confirmation (Failed) 16:18 - The Infinite Loop Problem 16:53 - Adding Agent Persistence and Memory 17:27 - Conversational Memory Still Causing Loops 17:55 - Gemini Acknowledges: "Too Complex" 18:25 - Manual Code Fix: String Matching Correction 19:05 - Understanding Nondeterministic LLM Responses 20:06 - Final Demo: Confirmation Pop-Up Working 20:50 - New Tag: "Last Human Approved Action" 21:22 - Reflections: Worth the Struggle 21:58 - Quiz: Testing Your Understanding 22:28 - Critical System Interaction Classification 23:02 - Next Topic Preview: Model Context Protocol (MCP) Network Intelligence - Real development, real challenges, real solutions for securing AI agents. #RedTeaming #AgenticAI #Azure #CloudSecurity #HumanInTheLoop #DenialOfWallet #CriticalInfrastructure #CSA #Cybersecurity #KKMookhey #NetworkIntelligence #AIRedTeam #AzureSecurity #RateLimiting #SecureAI