Grainger College of Engineering Seminars & Speakers

Refreshments will be provided.

Abstract: 
Modern AI security research has largely focused on model-centric threats such as adversarial examples, jailbreaks, backdoors, and alignment failures. Yet as AI systems evolve from standalone models into autonomous agents embedded in complex software stacks, security risks increasingly arise from system-level interactions rather than isolated model behavior. In this talk, I argue for a shift from model-level AI safety to a systems-oriented view of AI security that spans component, pipeline, interaction, and human-in-the-loop concerns, and connect it to established principles such as isolation, least privilege, and defense in depth. I illustrate this perspective through three recent works unified by a memory-centric defense paradigm: (1) treating key-value (KV) caches as a controllable boundary to limit adversarial context propagation in LLMs; (2) extending KV-cache defenses to vision-language models through dynamic token reweighting; and (3) designing a shadow memory architecture that protects LLM agents against persistent, multi-turn adversarial threats. Together, these works demonstrate that despite the growing complexity of AI systems, a broad class of security challenges can be addressed through unified, principled approaches rooted in systems thinking. I conclude by outlining an agenda centered on system-level threat modeling, architecture-aware defenses, and principled evaluation for securing the next generation of AI systems.

Bio:
Ting Wang is an Associate Professor and Empire Innovation Scholar in the Department of Computer Science at Stony Brook University. He received his Ph.D. in Computer Science from the Georgia Institute of Technology. His research centers on building security and privacy foundations for emerging technologies, with a current focus on securing AI-powered agentic systems. His work has appeared in premier computer security and AI venues, including IEEE S&P, CCS, USENIX Security, NeurIPS, ICLR, and CHI, and has been recognized with multiple best paper awards and industry adoption. His research is supported by NSF, DARPA, and industrial partners, including an NSF CAREER Award and an NSF CRII Award. He is also the recipient of two distinguished professorships.

Faculty Host: Gang Wang

Meeting ID: 831 7134 6416; Password: csillinois