Reproduce “ClawSentry: A Progressive Multi-Tier Security Monitor for Safeguarding Autonomous LLM Agents” 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

Kai Wang, Zeming Wei, BiaoJie Zeng, Chang Jin, An Wang, Xiaokun Luan, Zhixiao Lin, Jingjing Qu and 2 more

ClawSentry, developed by Shanghai Artificial Intelligence Laboratory, introduces a progressive multi-tier security monitor to safeguard autonomous LLM agents by mediating risks across their lifecycle and detecting bypass attempts. The system reduced contextual Attack Success Rate on SkillInject by 3.5–15.2x and confined ASR on SkillsSafety to 9.09–15.03%, while preserving clean task success at 98.7% and achieving an 86.2% reduction in model-backed security review calls.

Key insights

  • Agent security risks are "progressive," originating at various lifecycle stages and capable of re-manifesting through "capability-equivalent bypasses" that require stateful, cross-stage monitoring.
  • An architectural solution that decouples security policy from agent internals, utilizing an abstract protocol, allows for unified and comprehensive supervision across heterogeneous agent frameworks.
  • Efficiently funneling security checks through a progressive adjudication engine (deterministic L1, semantic L2, multi-turn L3) significantly reduces computational overhead by selectively escalating review to resource-intensive models.

Summary by alphaXiv.

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