Reproduce “LLM agents security duality: a comprehensive survey of self-security and empowered cybersecurity” 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

Yiwei Xu, Yong Zhuang, Xuanming Liu, Tian Zhang, Bowen Xiao, Xiaoyang Xu, Delong Jiang, Juan Wang and 1 more

The survey examines agent security as two connected problems: protecting tool-using agents and applying them to cyber offense and defense. It links recurring threats to six architectural mechanisms, including instruction-boundary collapse, persistent state hazards, and authority amplification through tools. A reported pop-up attack achieved an 86% average success rate and reduced task success by 47% in OS-World and VisualWebArena, but these figures are not universal agent rates.

Key insights

  • LLM agents present a "security duality," where enhancements in their self-security are critically linked to their effectiveness in empowering cybersecurity, fostering an ongoing "arms race."
  • Beyond inherent LLM flaws, agent-specific vulnerabilities emerge from dynamic interactions with environments, tools, and other agents, introducing new risks such as action hijacking and tool abuse.
  • Agents can transform cybersecurity by enabling proactive, intent-based defense and highly automated, sophisticated offensive capabilities, moving beyond traditional rule-based and manual processes.

Summary by alphaXiv.

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