📄 中文摘要
在运行 Claude Code 超过 200 小时后，作者建立了多代理系统，使两个 AI 能够在其休息时相互咨询。尽管认为设置稳固，但在运行自制的 10 项诊断脚本时，发现有两个项目出现了问题。这 10 项可能出错的事项源于真实的自主操作失败，包括缺少持久指令的 CLAUDE.md、未安装钩子导致自动检查无法运行，以及缺乏对危险命令的保护等。这些问题可能导致错误未被及时发现，从而影响系统的安全性和稳定性。
我对 Claude 代码设置进行了安全扫描——发现了什么
出处: I Ran a Safety Scan on My Claude Code Setup — Here's What I Found
发布: 2026年2月15日
📄 English Summary
I Ran a Safety Scan on My Claude Code Setup — Here's What I Found
After running Claude Code autonomously for over 200 hours, the author established a multi-agent system allowing two AIs to consult each other while he sleeps. Despite believing the setup was solid, running a self-created 10-item diagnostic script revealed two issues. These 10 potential failures stem from real incidents during autonomous operation, including the absence of a persistent CLAUDE.md for instructions, lack of installed hooks for automated checks, and insufficient protection against dangerous commands. Such issues could lead to unnoticed errors, impacting the system's safety and stability.