[1I5-GS-4b] Web intelligence

5:40 PM - 5:55 PM JST(8:40 AM - 8:55 AM UTC)

[1I5-GS-4b-01] Asymmetric Defense Against Web Scraping Attacks by Large Language Model-Driven Agents via Context Pressure and Resource Exhaustion

〇Mao Sonobe¹, Norifumi Watanabe¹ (1. Musashino University)

Keywords:

Large Language Models,Web Scraping,Prompt Injection,Honeypot,AI Security

The democratization of Large Language Models (LLMs) has lowered the technical barriers for web scraping and cyberattacks, rendering traditional rule-based defenses such as IP blocking ineffective against autonomous agents. We propose an active defense method designed to invert the cost asymmetry between attackers and defenders. By dynamically generating "Recursive Pseudo-Structured Data" (digital labyrinths) combined with "Structural Prompt Injection," the system forces attacking agents to exhaust their computational resources and financial budgets. Experimental results using GPT-4o show that the method increases attacker's economic cost by approximately 3,360 times and processing time by over 15 times compared to legitimate users. Furthermore, the structural injection successfully neutralized the attacker's system prompts with a 100% success rate, proving that defenders can establish an overwhelming asymmetric advantage in the AI era.