Title:
Verification of Frontier AI Models
Abstract:
Artificial intelligence (AI) systems are becoming increasingly capable of performing complex tasks, from answering questions to assisting with scientific research. The most advanced of these systems, often called frontier AI, offer enormous opportunities but also raise important questions about security, safety, and trust. How can we know whether an AI system behaves as promised? Could someone secretly modify it, introduce harmful capabilities, or bypass its safety protections?
This talk introduces AI verification, the process of checking claims about AI systems with trustworthy evidence rather than relying solely on developers’ reports. A central theme is understanding who might attempt to manipulate an AI system, what they can control, and how verification can expose such manipulation. We explore three main approaches: human-led inspections, software-based techniques, and hardware-based security mechanisms. Through practical examples, we examine what these approaches can verify today, which threats they can detect, and where their limitations lie.
Speaker:
Fatemeh (Saba) Ganji
Associate Professor, ECE Department, WPI
Bio:
Fatemeh (Saba) Ganji is an Associate Professor of Electrical and Computer Engineering at Worcester Polytechnic Institute (WPI), where she is also affiliated with the Cybersecurity program. Her research lies at the intersection of hardware security, cryptography, and machine learning, focusing on the security and trustworthiness of AI systems and computing hardware.
Host: Professor John McNeill