---
id: 20260423-T0-14
title: "研究提出计算机使用代理的人类引导危害恢复方案"
title_en: "Human-Guided Harm Recovery for Computer Agents"
url: https://ai.daily.yangsir.net/daily/20260423-T0-14
issue_date: 2026-04-23
publish_date: 2026-04-22T04:00:00.000Z
category: research
source_name: "arXiv cs.AI"
source_url: https://arxiv.org/abs/2604.18847
---

# 研究提出计算机使用代理的人类引导危害恢复方案

最新研究提出一种人类引导的危害恢复框架，专门解决LM代理在计算机系统执行操作时的潜在危害问题。该方案首次对事后危害补救进行了形式化定义，通过分层恢复策略和人类监督机制，有效解决了大规模预防失效后的补救难题。研究成果已发布在arXiv预印本平台，为构建更安全的AI代理系统提供了新思路。

## English Version

**Human-Guided Harm Recovery for Computer Agents**

A new study introduces a human-guided harm recovery framework for LM agents operating on computer systems. The research formally defines post-harm remediation as a previously neglected challenge, proposing layered recovery strategies with human oversight. Published on arXiv, this work offers a new approach to building safer AI agents by addressing failures in large-scale prevention systems.

---

**来源**：[arXiv cs.AI](https://arxiv.org/abs/2604.18847)

**详情页**：https://ai.daily.yangsir.net/daily/20260423-T0-14

---

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