LeanGuard adds Lean-verified policy checks for agent tool calls
Sreehari Sankar and 19 coauthors present LeanGuard, a runtime that checks every proposed tool call against a Lean4 policy before execution, using state predicates and scoped past-time temporal operators. The authors say evaluation fails closed, the decision procedure is proved equivalent to LeanLTL semantics, and proof-guided repair can find an admissible edit without model involvement. In experiments on Tau-Bench with five Qwen models, DTap and a telecom task, the paper reports no policy-forbidden calls under LeanGuard, reduced direct prompt-injection success from 57.9% to 5.9% on DTap, and unchanged benign-task success in telecom.
The paper offers a formally specified, fail-closed way to constrain tool-using agents before actions reach an environment, alongside evaluations on multiple benchmarks and a proof-guided repair mechanism.
Computer Science > Artificial Intelligence
arXiv:2611.02332 (cs)
[Submitted on 2 Nov 2026]
Title:LEANGUARD: Lean-Verified Guardrails for Tool-Using Agents
Authors:Sreehari Sankar, Debargha Ganguly, Vikash Singh, Mohsen Hariri, Weicong Chen, Yanyan Zhang, Shouren Wang, Wang Yang, Chaoda Song, Alexander Nemecek, Osama Zafar, Idan Davidovich, Yuting Shao, Kai Ye, Biyao Zhang, Xinpeng Li, Joe Stubbs, Erman Ayday, Jing Ma, Vipin Chaudhary
View a PDF of the paper titled LEANGUARD: Lean-Verified Guardrails for Tool-Using Agents, by Sreehari Sankar and 19 other authors