Microsoft Chief AI Officer: How to Fix Broken LLMs | Dr. Paul Rodrigues

Dr. Paul Rodrigues, Chief AI Officer for Microsoft's National Security Group, reveals the hidden layers of risk within large language models and how to defend them. In this comprehensive briefing, learn why the technology industry is facing unprecedented security threats due to releasing highly advanced AI models before we completely understand how their inner layers interact. Dr. Rodrigues explains the "AI Onion", a structural model mapping vulnerabilities from core weight alignment and reinforcement learning up to user experience applications. Discover the true engineering challenge of locating hidden objectives inside neural networks, identifying dataset bias introduced during reinforcement learning with human feedback (RLHF), and establishing bulletproof defense infrastructures. This talk outlines the operational blueprints used by the Microsoft AI Red Team to holistically stress-test systems, bypass guardrails safely, and implement point-and-click Retrieval Augmented Generation (RAG) frameworks like GraphRAG to keep data securely grounded.⏱️ Timestamps 0:00 - Intro: Dr. Paul Rodrigues, Chief AI Officer 0:35 - Mapping Vulnerabilities Across LLM Layers 2:23 - The History and Evolution of NLP Models 7:55 - AI Alignment and Human Preference Risks 11:37 - Uncovering Hidden Objectives and Ethics 14:02 - Platform Security: Guardrails, RAG, & Red Teaming 21:29 - Q&A: Automation Paradox and Dataset Diversity Key Takeaways The Hidden Core: Inspecting an artificial intelligence model for hidden motivations or malicious alignment still remains an incredibly labor-intensive, manual process involving interactive interrogation. Dataset Value Bias: When organizations establish specialized annotation guidelines for human preference pools, they inherently bake explicit cultural, organizational, or ethical perspectives into the model's downstream operations. Holistic Defense: Security cannot focus solely on the base neural network; safety teams must audit the entire integrated application infrastructure—including guardrails, grounding rules, meta-prompting, and systemic data logging. #Cybersecurity #generativeai #largelanguagemodels #aisecurity #techbriefing #datasovereignty