CV
Research Interests
AI safety and adversarial robustness of large language models: red-teaming of LLM and computer-use agents, indirect prompt injection, jailbreaking, and safety alignment.
Education
Simon Fraser University, Burnaby, Canada — Ph.D. in Computing Science, Sep 2022 – Present
- Advised by Prof. Ke Wang. GPA: 4.21/4.33
- Awards: Special Graduate Entrance Scholarship, Graduate Fellowship (×6)
Shenzhen University, Shenzhen, China — B.E. with Honors in Artificial Intelligence, Sep 2018 – Jun 2022
- GPA: 3.96/4.5 (Rank: 4/103)
- Awards: Outstanding Graduate, Outstanding Bachelor Thesis, Star of SZU Scholarship, MCM Meritorious Winner
Research Experience
RBC Borealis (Borealis AI), Vancouver, Canada — Machine Learning Researcher (Intern), Jan 2026 – May 2026
Project: Computer-Use Agent Safety
- Introduced multi-step indirect prompt injection, a new attack class that decomposes an adversarial goal into innocuous sub-steps spread across a chain of web pages.
- Built an automatic LLM-driven pipeline for adversarial goal decomposition.
- Released the CUA safety benchmark StepJack, and evaluated six state-of-the-art computer-use agents.
- Raised attack success rate on GPT-5.4-mini from 41.7% at single-step to 72.9% at three-step.
- Research paper submitted, with patent filed and code open-sourced.
Publications & Preprints
- Zhuoxin Zhan, A. Rafiey, A. Ma, L. Pishdad, L. El Asri. StepJack: Benchmarking Computer-Use Agent Safety Against Multi-Step Indirect Prompt Injection. Preprint, arXiv 2026. paper · code
- Zhuoxin Zhan, K. Wang, P. Xiong, L. Li. Benign Prompts Can Jailbreak Large Language Models. Under submission, 2026.
- Zhuoxin Zhan, K. Wang, P. Xiong. Accelerating Adversarial Training on Under-Utilized GPU. IJCAI 2025. paper · code
- Zhuoxin Zhan, M. He, W. Pan, Z. Ming. TransRec++: Translation-Based Sequential Recommendation with Heterogeneous Feedback. Frontiers of Computer Science, 2022. paper
- Zhuoxin Zhan, L. Zhong, J. Lin, W. Pan, Z. Ming. Sequence-Aware Similarity Learning for Next-Item Recommendation. The Journal of Supercomputing, 2021.
- Y. Ni, Zhuoxin Zhan, W. Pan, Z. Ming. Asymmetric Pairwise Preference Learning for Heterogeneous One-Class Collaborative Filtering. ICONIP 2020.
- J. Lin, Zhuoxin Zhan, W. Pan, Z. Ming. Single-Behavior Sequential Recommendation. Book chapter, Intelligent Recommendation Technology, Tsinghua University Press, 2022 (in Chinese). book
Services
Conference Reviewer — NeurIPS '26, ICML '26, EMNLP '26, WWW '26, WSDM '25, CIKM '25/'24, KDD '24, BigData '24
Student Chair, Artificial Intelligence Club, Shenzhen University — Sep 2019 – Jul 2021
- Collaborated with sponsors (Tencent, IBM, Baidu) to organize events such as IBM summer camps and Baidu seminars.
Technical Skills
- Languages: Python, C/C++, Java, SQL
- ML / AI: PyTorch, HuggingFace Transformers, vLLM, scikit-learn, NumPy, Pandas, Weights & Biases