WebSentinel: Detecting and localizing prompt injection attacks against web agents

The researchers propose WebSentinel, a two-step method for detecting and localizing prompt injection attacks in web pages, where such attacks manipulate web page content to induce web agents to perform attacker-specified tasks. The first step extracts potentially contaminated segments of interest from the web page, and the second step checks whether each segment is consistent with the web page content as context, thereby determining contamination. The authors state that on their collected datasets of multiple contaminated and clean web pages, WebSentinel substantially outperforms baseline detection methods. Code is open-sourced at https://github.com/wxl-lxw/WebSentinel。