---
id: 20260310-T0-06
title: "Traversal-as-Policy：LLM 智能体安全框架"
title_en: "Traversal-as-Policy: New LLM Agent Security Framework"
url: https://ai.daily.yangsir.net/daily/20260310-T0-06
issue_date: 2026-03-10
publish_date: 2026-03-09T04:00:00.000Z
source_name: "arXiv cs.LG (ML)"
source_url: https://arxiv.org/abs/2603.05517
---

# Traversal-as-Policy：LLM 智能体安全框架

Traversal-as-Policy 是一种新的智能体安全框架，通过提取 OpenHands 执行日志生成可验证的门控行为树。该方法将长期策略显式化，避免传统 LLM 智能体的隐式策略导致的安全问题。实验表明，该框架在保持任务完成率的同时，将越界行为减少 67%。代码已在 GitHub 开源。

## English Version

**Traversal-as-Policy: New LLM Agent Security Framework**

Traversal-as-Policy is a new security framework that extracts verifiable guard behavior trees from OpenHands execution logs. It makes long-term policies explicit, avoiding security issues from implicit policies in traditional LLM agents. Experiments show it reduces out-of-bound behaviors by 67% while maintaining task completion rates. The code is open-sourced on GitHub.

---

**来源**：[arXiv cs.LG (ML)](https://arxiv.org/abs/2603.05517)

**详情页**：https://ai.daily.yangsir.net/daily/20260310-T0-06

---

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