---
id: 20260512-T0-06
title: "大模型安全审计新方法：统一图表示"
title_en: "New LLM Security Audit Method: Unified Graph"
url: https://ai.daily.yangsir.net/daily/20260512-T0-06
issue_date: 2026-05-12
publish_date: 2026-05-11T04:00:00.000Z
category: research
source_name: "arXiv cs.AI"
source_url: https://arxiv.org/abs/2605.06812
---

# 大模型安全审计新方法：统一图表示

研究提出基于图的LLM代理安全审计框架，将动态工具调用、内存管理等复杂行为转化为可验证的图结构。该框架能自动检测代理系统中的语义漏洞，在5个基准测试中识别出12类高风险缺陷。开发者可通过可视化界面追溯代理决策链，大幅提升企业级AI系统的安全性。

## English Version

**New LLM Security Audit Method: Unified Graph**

Researchers propose a graph-based LLM agent security audit framework that converts complex behaviors like tool invocation into verifiable graph structures. It automatically detects semantic vulnerabilities and identifies 12 high-risk defects across 5 benchmarks.

---

**来源**：[arXiv cs.AI](https://arxiv.org/abs/2605.06812)

**详情页**：https://ai.daily.yangsir.net/daily/20260512-T0-06

---

*智语观潮 · Daily — https://ai.daily.yangsir.net/llms.txt*