ACL 2026

Benchmarking Web Agent Safety under E-commerce Deceptive Interfaces

Zijing Shi1, Meng Fang2, Ling Chen1
1University of Technology Sydney  ยท  2University of Liverpool
Paper Code
WebDecept overview: end-to-end shopping task, intervention layer, deceptive patterns, and findings.
Illustration generated with ChatGPT Images 2.

We introduce WebDecept, a lightweight plugin framework that injects controlled deceptive interface patterns into a reproducible e-commerce environment, and use it to evaluate the safety of multimodal web agents in end-to-end shopping tasks.

BibTeX
@inproceedings{shi2026webdecept,
  title     = {Benchmarking Web Agent Safety under E-commerce Deceptive Interfaces},
  author    = {Shi, Zijing and Fang, Meng and Chen, Ling},
  booktitle = {Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics},
  year      = {2026}
}
Acknowledgments

WebDecept is built on top of VisualWebArena; we thank the authors for releasing the valuable environment.