Reproduce “Genesis: Evolving Attack Strategies for LLM Web Agent Red-Teaming” 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

Zheng Zhang, Jiarui He, Yuchen Cai, Deheng Ye, Peilin Zhao, Ruili Feng, Hao Wang

Genesis introduces an agentic red-teaming framework that autonomously discovers and evolves attack strategies for large language model (LLM) web agents, leveraging a genetic algorithm and hybrid strategy representation. The system significantly outperforms existing baselines, achieving an average attack success rate (ASR) of 53.0% against the SeeAct agent with GPT-4o, and demonstrates high transferability of discovered strategies across different LLM backends.

Key insights

  • An agentic framework with iterative learning is crucial for effectively red-teaming LLM web agents, enabling autonomous discovery and evolution of attack strategies.
  • Combining natural language descriptions and executable code for strategy representation enhances the robustness and comprehensiveness of generated attacks.
  • The discovered attack strategies are transferable across different backend LLMs, indicating they exploit generalizable vulnerabilities in web agents rather than model-specific artifacts.

Summary by alphaXiv.

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