This research introduces a novel Denial-of-Service attack technique specifically designed to exhaust computational resources in LLMs under black-box settings.

• Develops an automated algorithm that generates resource-intensive prompts to overwhelm LLM systems
• Demonstrates successful attacks against multiple commercial LLM services
• Identifies critical security vulnerabilities in current LLM deployment architectures
• Proposes potential defense mechanisms for LLM providers

This work highlights significant security concerns for enterprise LLM deployments, showing how attackers might disrupt AI services without requiring internal system knowledge. Understanding these attack vectors is essential for implementing robust security measures in production language model systems.

Crabs: Consuming Resource via Auto-generation for LLM-DoS Attack under Black-box Settings