Reproduce “From Shield to Target: Denial-of-Service Attacks on LLM-Based Agent Guardrails” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Yuguang Zhou, Xunguang Wang, Pingchuan Ma, Zhantong Xue, Zhaoyu Wang, Shuai Wang

A new reasoning-extension Denial-of-Service (DoS) attack was discovered against LLM-based agent guardrails, exploiting their structured reasoning capabilities to force prolonged, computationally expensive analytical loops. This attack achieved up to 63x token amplification on various guardrail models and severely degraded performance in real-world agent deployments, demonstrating its effectiveness compared to prior LLM DoS methods.

Key insights

  • LLM guardrails' inherent structured reasoning and instruction-following fidelity, while designed for safety, can be exploited as a vulnerability for DoS attacks.
  • The attack leverages guardrails' mechanical template-filling behavior, characterized by 'attention cycling' and 'entropy collapse,' rather than genuine complex reasoning.
  • A 'safety-availability dilemma' arises, forcing systems to choose between denying service or risking unreviewed, potentially unsafe actions when guardrails are exhausted.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.