Qiusi Zhan | 詹秋思

I am a Ph.D. student at the Univerity of Illinois Urbana Champaign (UIUC), advised by Prof. Daniel Kang. Previously, I obtained my master's degree from UIUC, advised by Prof. Heng Ji, and completed my bachelor's degree from Peking University, advised by Prof. Sujian Li.

Email  /  CV  /  Google Scholar  /  Github

My research focuses on developing safe (multimodal) Large Language Models (LLMs) and LLM agents for real-world deployment, with an emphasis on identifying and mitigating safety vulnerabilities. I have studied a wide range of safety risks in LLMs and LLM agents, including fine-tuning vulnerabilities, indirect prompt injection attacks, multimodal RAG knowledge poisoning, and backdoor attacks. On the mitigation side, I have explored reinforcement learning approaches for LLM-based agents to enhance their safety without compromising utility.

We propose SafeSearch, the first safety alignment framework for search agents that enhances safety without compromising utility.

We introduce BEAT, the first framework to demonstrate visual backdoor attacks on multimodal large language model (MLLM) based embodied agents.

We test eight different defenses of Indirect Prompt Injection attacks and demonstrate that they are vulnerable to adaptive attacks.

We propose MM-PoisonRAG, a novel knowledge poisoning attack framework for multimodal RAG.

We benchmark IPI attacks in tool-integrated LLM agents and show that most of the agents are vulnerable to such attacks.

We demonstrate that fine-tuning GPT-4 with just 340 examples can subvert its RLHF protections, underscoring the critical need for enhanced security measures.

Reviewer for ARR, NeurIPS, and ICLR.
Recognized as “Top Reviewer” at NeurIPS 2025.

This website is based on a template created by Jon Barron.
Last updated: Oct 30, 2025