WebCloak: Characterizing and Mitigating Threats From LLM-Driven Web Agents as Intelligent Scrapers

The rise of web agents powered by large language models (LLMs) is reshaping the landscape of human-computer interaction, enabling users to automate complex web tasks with natural language commands. However, this progress introduces significant, yet largely unexplored security concerns: adversaries can employ such web agents to conduct largescale web scraping, particularly of visual content. This paper presents the first systematic characterization of the danger represented by such LLM-driven web agents as intelligent scrapers. We develop LLMCrawlBench, a large test set of 237 extracted real-world webpages (10,895 images) from 50 popular high-traffic websites in 5 critical categories, designed specifically for adversarial image extraction evaluation. Our metrics across over 32 scraper implementations, including LLMto-Script (L2S), LLM-Native crawlers (LNC), and LLM-based web agents (LWA), demonstrate that while some tools exhibit working issues, advanced LLM-powered frameworks lower the bar for effective scraping.

Such new agent-as-attacker threats motivate us to introduce WebCloak, an effective, lightweight defense that specifically targets the main weakness of LLM crawler agents' fundamental "Parse-then-Interpret" mechanism. Our key idea is dual-layered: (1) Dynamic Structural Obfuscation, which not only randomizes structural cues but also restores visual content client-side using non-traditional methods less amenable to direct LLM exploitation, and (2) Optimized Semantic Labyrinth to mislead the central LLM interpretation of the agent through added harmless-yetmisleading contextual clues, all while not sacrificing visual quality for legitimate users. Our evaluations demonstrate that WebCloak significantly reduces scraping recall rates from 88.7% to 0% against leading LLM-driven scraping agents, offering a robust and practical countermeasure.

Artist Images News Events Stock Data Travel Data