Dual-Modality Multi-Stage Adversarial Safety Training: Robustifying Multimodal Web Agents Against Cross-Modal Attacks
Accepted to Conference on Empirical Methods in Natural Language Processing (EMNLP), 2026
DMAST combines imitation learning, oracle-guided supervised fine-tuning, and adversarial reinforcement learning with GRPO self-play to jointly train a web agent and an attacker. The framework improves robustness to cross-modal attacks and task completion on unseen tasks.
Authors: Haoyu Liu, Dingcheng Li, Lukas Rutishauser, Zeyu Zheng