Multimodal Large Language Model (MLLM)-based web agents are increasingly used for visual browser automation, providing practical and high-precision solutions. But according to a new arXiv paper, these agents also inherently expand the attack surface, opening the door to novel vision-based attacks that have not been seen before.

The paper introduces MIRAGE, a stealthy visual prompt injection technique designed for vulnerability detection in web agents. Rather than relying on text-based manipulation, MIRAGE embeds malicious instructions directly into visual inputs, allowing an attacker to influence agent behavior without being easily noticed.

Because the abstract is truncated, the full details of the attack methodology and evaluation are not available in the source. Still, the core claim is clear: as web agents become more capable of interpreting visual information, they also become more exposed to attacks that exploit that very capability. The authors position MIRAGE as both a demonstration of the risk and a tool for identifying vulnerabilities before they can be exploited in the wild.