arXiv AI

Specification-based Security Evaluation of Real-World Agents Tool-Calling under Long-Horizon Context Tasks

The paper introduces RAS‑Eval, a specification‑based security evaluation framework for large language model agents that uses Linear Temporal Logic (LTL) to define utility and security objectives over tool‑call trajectories and dynamic environment observations. Unlike traditional LLM‑as‑a‑Judge or Agent‑as‑a‑Judge approaches, RAS‑Eval derives deterministic verdicts via a model checker, reducing evaluator uncertainty and enabling reliable assessment over long‑horizon tasks. The authors also present a multi‑agent dataset pipeline and a benchmark of 200 normal and 800 attack tasks covering indirect prompt injection, tool poisoning, skill code injection, and tool code injection. whyItMatters":"RAS‑Eval addresses key limitations of existing security evaluations—such as evaluator hallucination and missed transient violations—by providing deterministic, environment‑aware verdicts that are crucial for safely deploying LLM agents in unattended, long‑horizon scenarios."

arXiv AI
Sep 17

Symbolic Temporal Supervision of LLM Agents Using Contracts

ContrAgent is a contract‑based framework that provides symbolic temporal supervision for large language model agents. It records an agent’s tool‑call sequence as a trace of checkable predicates and formalizes desired behaviors with assume‑guarantee contracts expressed in linear temporal logic over finite traces (LTLf). Each contract is compiled into a deterministic finite automaton that both gates actions online and evaluates recorded traces offline, enabling deterministic, reproducible verdicts and significantly lower per‑call latency compared to existing LLM‑judge and rule‑based guardrail baselines.

By Yifeng Xiao, Pierluigi Nuzzo
arXiv AI
Jun 2

SeClaw: Spec-Driven Security Task Synthesis for Evaluating Autonomous Agents

arXiv:2606. 02302v1 Announce Type: cross Abstract: Autonomous LLM agents increasingly operate in stateful environments where they access tools, files, memory, and external services.

By Hao Cheng, Changtao Miao, Tianle Song, Yin Wu, He Liu, Erjia Xiao, Junchi Chen, Xiaoyu Shi, Yichi Wang, Jing Yang, Taowen Wang, Jinhao Duan, Mengshu Sun, Peiyan Dong, Xuan Shen, Yang Cao, Renjing Xu, Kaidi Xu, Jindong Gu, Bo Zhang, Jize Zhang, Chenhao Lin, Philip Torr, Chao Shen
arXiv AI
Jun 18

SafeClawBench: Separating Semantic, Audit-Evidence, and Sandbox Harm in Tool-Using LLM Agents

arXiv:2606. 18356v1 Announce Type: cross Abstract: Tool-using language-model agents introduce security failures that go beyond unsafe text: they can disclose protected objects, write persistent memory, send messages, modify databases, or trigger harmful code and tool effects.

By Yuchuan Tian, Mengyu Zheng, Haocheng Mei, Ye Yuan, Chao Xu, Xinghao Chen, Hanting Chen, Yu Wang