EN / KO

Activities

Seminars

Home Activities Seminars

FIELD
AI and Natural Sciences
DATE
Nov 25 (Wed), 2026
TIME
14:00 ~ 16:00
PLACE
7323
SPEAKER
Yujin Choi
HOST
Park, Jinseong
INSTITUTE
Sejong University
TITLE
Trustworthy LLM Agents: Privacy, Safety, and Reliability in Long-Horizon Interactions
ABSTRACT
Recent advances in large language models (LLMs), including long-context processing, in-context learning, and retrieval-augmented generation (RAG), have enabled the development of LLM agents that can perform complex tasks through reasoning, planning, and tool use. By incorporating long-term memory, these systems are evolving into long-lived, personalized agents capable of learning from past interactions and adapting to individual users. However, as agents operate over extended periods and accumulate information, new challenges arise concerning their privacy, safety, and reliability. In this talk, we will first introduce the fundamental architecture of LLM agents and discuss privacy risks associated with RAG systems. In particular, we will examine membership inference attacks (MIAs), which aim to determine whether specific data are included in a model's training data or an external retrieval database. We will also introduce our previous research on detecting and defending against membership inference attacks in RAG systems. Next, we will explore the challenges of ensuring trustworthiness in long-horizon LLM agents. One important aspect is reliability. As agents continuously accumulate and update their memories, outdated knowledge, memory compression, and improper revisions may lead to increasingly unreliable behavior. We will discuss these potential challenges and introduce recent research on long-term memory management. Additionally, we will examine potential safety and privacy risks associated with personalized agents. Over time, accumulated interactions may enable agents to infer and disclose sensitive information that users never explicitly shared, while personalized memories may also affect their safety. We will discuss these emerging concerns and highlight open questions and possible research directions toward trustworthy long-horizon LLM agents.
FILE