Laizhen Li

Laizhen Li (李来真)

Master's Student in Computer Technology @ UCAS

lilaizhencs@gmail.com GitHub Google Scholar CV (PDF)

About Me

I am a master’s student in Computer Technology at the University of Chinese Academy of Sciences (UCAS). I conduct research at the Research Center for Cloud Computing, Shenzhen Institutes of Advanced Technology (SIAT), Chinese Academy of Sciences, under the supervision of Xitong Gao. I received my B.Eng. in Software Engineering from Henan University of Engineering in 2023.

I work on large language models (LLMs) and AI agents, with a focus on safety, security, and efficiency.

Selected Works

StreamGuard workflow: self-review of partial generations followed by continuation or safety prompt insertion.

LLM Safety - StreamGuard

Paper

Designed and implemented StreamGuard, a lightweight, plug-and-play decoding-time defense against jailbreak attacks. StreamGuard self-reviews partial generations and inserts safety prompts when unsafe content is detected, without fine-tuning. Across six model-attack settings, it achieved a mean attack success rate of 1.7%, compared with 20.3% for the strongest defense baseline and 60.3% without defense.

A2M framework: metadata optimization in the Attraction phase and trace-guided payload refinement in the Manipulation phase.

Agent Security - A2M

Paper

Designed and implemented A2M (Attraction to Manipulation), a two-stage black-box attack that optimizes tool metadata to increase selection likelihood and then refines return payloads using execution traces to manipulate downstream reasoning. On LiveMCPBench, A2M achieved a mean ASR of 74.4% across information exfiltration, unauthorized environment modification, and reasoning derailment objectives, and increased token consumption by 32.4x on a separate token-wasting objective. The optimized attacks also transferred effectively to unseen models.

Growing Harness framework: a minimal scaffold grows through failure tracing, localized code repairs, and held-out evaluation with rollback.

Agent Efficiency - Growing Harness

Paper

Designed and implemented Growing Harness, a failure-guided framework that starts from a strategy-free scaffold and incrementally grows into an executable agent harness through function-level execution traces and localized code edits. This shifts recurring control logic from LLM inference to reusable program code. Across six benchmark-model settings, it achieved a mean task success rate of 42.2%, compared with 23.6% for a ReAct-style Tool-Calling baseline, while reducing LLM calls by 76.0-91.8% and inference cost by 74.4-98.6%.

Publications

  1. A2M: Trace-Optimized Agent Hijacking in the MCP Ecosystem
    Laizhen Li, Xuan Wang, Peicheng Zhao, Juanjuan Zhao, Kejiang Ye, Cheng-zhong Xu, and Xitong Gao.
    AACL-IJCNLP 2026 (Main Conference, Oral Presentation). Paper

  2. StreamGuard: A Streaming-based Defense Against Jailbreaking Attacks in Large Language Models
    Laizhen Li, Xitong Gao, Xuan Wang, Juanjuan Zhao, and Kejiang Ye.
    ICONIP 2025. Paper

  3. JADE: Jailbreak-Aware Dynamic Defense Enhancer for LLMs via Reinforcement Learning
    Peicheng Zhao, Laizhen Li, Xitong Gao, Juanjuan Zhao, and Kejiang Ye.
    IEEE BigData 2025. Paper

Manuscripts

  1. Grow the Harness, Not the Context: From Strategy-Free Scaffolds to Reusable Specialist Agents
    Laizhen Li, Jiarui Li, Juanjuan Zhao, Kejiang Ye, Ye Li, Cheng-zhong Xu, and Xitong Gao. Paper