Title: Building Faithful and Secure LLM Agents

Speaker: Prof. Hwanhee Lee @ Chung-Ang University

Time : 11:00 - 12:00, Jan 19th, 2026

Location: Hybrid

In-person: 26310

Online: https://hli.skku.edu/InvitedTalk260119 

Language: English speech & English slides

Abstract:

The rapid advancement of Large Language Models (LLMs) has unlocked remarkable capabilities, increasingly empowering them to function as autonomous agents. However, their deployment in high-stakes applications is severely hampered by persistent issues of insecurity and unreliability. In this talk, I address these challenges through the dual lenses of “Security” and “Faithfulness” in LLM agents. First, I will introduce the concept of “contextual security,” examining vulnerabilities where LLMs fail to adhere to user-defined security policies. This includes presenting the first large-scale benchmark for policy preservation against indirect attacks and analyzing a novel attack vector, ChatInject, that exploits agents' native chat templates. Second, I will explore “contextual sensitivity” as a fundamental dimension of faithfulness. I will demonstrate how LLMs exhibit unexpected biases induced by the language of retrieved information and the sentiment of user personas, leading to degraded performance and unfair outcomes. Ultimately, this talk provides a comprehensive diagnosis of critical failures in modern LLMs and offers a clear path toward building the secure, faithful, and reliable AI agents of the future.

Bio:

Hwanhee Lee is an Assistant Professor in the Department of Artificial Intelligence at Chung-Ang University. His primary research field is Natural Language Processing (NLP), where he focuses on the trustworthiness and safety of language models. He was recognized as an Outstanding Area Chair at EMNLP 2024. Prior to joining Chung-Ang University in March 2023, he was a senior researcher at the Automated System Research Institute. He received his Ph.D. (2022) and B.S. (2017) in Electrical and Computer Engineering from Seoul National University.