Publications

StepJack: Benchmarking Computer-Use Agent Safety Against Multi-Step Indirect Prompt Injection

Zhuoxin Zhan, Akbar Rafiey, Avery Ma, Leila Pishdad, Layla El Asri

Preprint, arXiv 2026

  • ▸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.
PaperCode

Benign Prompts Can Jailbreak Large Language Models

Zhuoxin Zhan, Ke Wang, Pulei Xiong, Linyi Li

Under submission 2026

  • ▸Demonstrated that benign prompts can trigger harmful responses in safety aligned LLMs.
  • ▸Proposed Benign Prompt Attack Paradigm for boosting existing jailbreak attacks, lifting attack success rates by 35–45%.
  • ▸Introduced adaptive safety alignment by incorporating (Benign prompt, Harmful response) conversations to mitigate the new attacks.

Accelerating Adversarial Training on Under-Utilized GPU

Zhuoxin Zhan, Ke Wang, Pulei Xiong

Proceedings of the 34th International Joint Conference on Artificial Intelligence (IJCAI) 2025

  • ▸Proposed AttackRider, consisting of a subset selection method and an example packing method, to accelerate adversarial training.
  • ▸Achieved 2–4× speedup at comparable robust accuracy across NN, CNN, and Transformer models on image and tabular tasks.
PaperCode

TransRec++: Translation-Based Sequential Recommendation with Heterogeneous Feedback

Zhuoxin Zhan, Mingkai He, Weike Pan, Zhong Ming

Frontiers of Computer Science 2022

Paper

Single-Behavior Sequential Recommendation

Jing Lin, Zhuoxin Zhan, Weike Pan, Zhong Ming

Intelligent Recommendation Technology (Book Chapter, in Chinese) 2022

Paper

Sequence-Aware Similarity Learning for Next-Item Recommendation

Zhuoxin Zhan, Liulan Zhong, Jing Lin, Weike Pan, Zhong Ming

The Journal of Supercomputing 2021

Asymmetric Pairwise Preference Learning for Heterogeneous One-Class Collaborative Filtering

Yongxin Ni, Zhuoxin Zhan, Weike Pan, Zhong Ming

Proceedings of the 27th International Conference on Neural Information Processing (ICONIP) 2020