MIRAGE: Stealthy Visual Prompt Injection for Vulnerability Detection in Web Agents

The paper introduces MIRAGE, a visual indirect prompt injection framework designed for vulnerability detection in Multimodal Large Language Model (MLLM)-based web agents. It utilizes diffusion models to create adversarial images that are visually benign and confined to specific spatial regions controlled by the attacker, employing a robust optimization technique that enhances attack effectiveness. This research is significant for practitioners as it highlights new vulnerabilities in web agents and provides a method for evaluating their security under realistic constraints, emphasizing the need for improved defenses against such stealthy attacks.