Dual-Modality Multi-Stage Adversarial Safety Training: Robustifying Multimodal Web Agents Against Cross-Modal Attacks

DMAST is a three-stage adversarial safety training pipeline for multimodal web agents that jointly reasons over screenshots and accessibility trees. It targets cross-modal DOM injection attacks, substantially reducing adversarial risk while improving efficiency on out-of-distribution MiniWob++ tasks.

• WebSentinel: Detecting and Localizing Prompt Injection Attacks for Web Agents
  February 3, 2026 · arXiv
• The Blind Spot of Agent Safety: How Benign User Instructions Expose Critical Vulnerabilities in Computer-Use Agents
  April 12, 2026 · arXiv
• Preference Redirection via Attention Concentration: An Attack on Computer Use Agents
  April 9, 2026 · arXiv
• A Survey on the Safety and Security Threats of Computer-Using Agents: JARVIS or Ultron?
  May 16, 2025 · arXiv
• WebSP-Eval: Evaluating Web Agents on Website Security and Privacy Tasks
  April 7, 2026 · arXiv
• WebTrap Park: An Automated Platform for Systematic Security Evaluation of Web Agents
  January 13, 2026 · arXiv