Reproduce “Cumulative and Audited Guardrail Tool Evolution for LLM Agent Safety” 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

Xiao Yang, Aishan Liu, Siyuan Liang, Zonghao Ying, Le Wang, Jiakai Wang, Xianglong Liu, Dacheng Tao

EVOLVE develops executable guardrail tools for agent risks, reusing and combining earlier tools and auditing them before library admission. Risk analysis identifies known or emerging risks, then tools are tested and repaired or discarded before reuse. On the 571-record RJudge benchmark, DeepSeek-V3.2 with EVOLVE reaches 0.9065 F1 for unsafe interactions, versus 0.8318 for ShieldAgent. The evidence is benchmark-based and does not establish safe open-ended deployment; tool auditing uses an LLM judge, not a formal security proof.

Summary by alphaXiv.

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