---
id: 20260807-T0-07
title: "研究揭示：人类在4万次游戏运行中遗漏1/3的AI代理威胁"
title_en: "Humans Missed 1 in 3 Threats When Approving AI Agent Commands in Study"
url: https://ai.daily.yangsir.net/daily/20260807-T0-07
issue_date: 2026-08-07
publish_date: 2026-08-06T11:58:07.000Z
category: research
source_name: "HN AI 精选"
source_url: https://scalex.dev/blog/ai-agent-permissions-stats/
---

# 研究揭示：人类在4万次游戏运行中遗漏1/3的AI代理威胁

一项针对AI代理安全的研究发现，人类在批准AI代理命令时，大约会遗漏三分之一的潜在威胁。研究团队进行了4万次游戏运行测试，统计用户在审查AI代理操作时的判断准确率。结果显示，人类监督者面对大量需要审批的命令时，注意力下降明显。该研究对当前"人类在环"的AI安全模式提出警示，建议开发更有效的自动化威胁检测机制。

## English Version

**Humans Missed 1 in 3 Threats When Approving AI Agent Commands in Study**

A study on AI agent safety found humans miss approximately one-third of potential threats when approving agent commands. Researchers ran 40,000 game sessions to measure user accuracy in reviewing AI agent actions. Results show human attention drops significantly with high approval volume. The findings challenge the 'human-in-the-loop' safety model and suggest developing better automated threat detection.

---

**来源**：[HN AI 精选](https://scalex.dev/blog/ai-agent-permissions-stats/)

**详情页**：https://ai.daily.yangsir.net/daily/20260807-T0-07

---

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