Data Poisoning - The Hidden Risk Shaping AI
This episode explores data poisoning and its growing impact on AI systems, from model backdoors to agent memory risk. Ioana and Chris chat with Microsoft's Giorgio Severi about how adversaries manipulate data, why these attacks are hard to detect, and what it takes to build layered defenses that keep AI systems reliable, safe, and trustworthy.
What You Will Learn:
Understand what AI red teaming is and why it’s critical for safe AI deployment
Learn how data and model poisoning can subtly influence AI behavior over time
Explore why AI systems can fail silently (e.g., backdoors and hidden triggers)
Discover the importance of layered security (“defense in depth”) in AI systems
Gain insight into new risks in AI agents, especially around memory and persistence
Get practical guidance on how to design and test more trustworthy AI systems
Guest bio
Giorgio Severi is a Senior AI Safety Researcher at Microsoft, where he works on the AI Red Team to assess the security and safety of large, multimodal, and agentic AI systems. His research focuses on adversarial machine learning, particularly risks related to poisoning and long-term memory. Before joining Microsoft, Giorgio completed his PhD at Northeastern University and has also worked at Sapienza University of Rome. His work has been recognized with a prize at the International Nasa SpaceApps Challenge 2015 for the Cropp project, which helps farmers monitor their lands.
