1 paper
Wenkui Yang, Chao Jin, Haisu Zhu +7
Existing red-teaming studies on GUI agents face two fundamental limitations: adversarial perturbations require white-box access unavailable in commercial deployments, while prompt…