reddit · SideProject · 09-14 23:50 65 分 ← 返回列表

Five weeks of letting an AI maintain my repo. It cut approval spam 99%, then allowed "drop the production database cluster"

A fake "drop the production database cluster" scored 95 and my own guard allowed it. No policy matched. That was Sep 8 and it's the most useful thing that's happened to this project. Quick background. DashClaw is the approval layer I built so agents running unattended have to ask before they do something destructive. An AI maintains the repo under a constitution I hold and writes a public log of what it did. Last update was five weeks ago. Back in August I turned every polic

五维评分

付费意愿

需求真实度

个人可落地

竞争空间

新鲜度

AI 分析

痛点：AI代理审批疲劳会导致用户关闭安全策略，而简单阈值调优又会放过“删除生产数据库”等高危操作，需要可自托管的智能审批与不可覆盖的高危操作下限。

目标用户：运行无人值守AI编码/运维代理的独立开发者、小团队和平台工程团队。

证据：1,759 approval interruptions in seven days. About 251 a day, against 2 blocks. ... A fake "drop the production database cluster" scored 95 and my own guard allowed it.

变现：按代理数或席位的SaaS订阅，例如每代理每月20-50美元；也可提供企业自托管授权、审计日志与合规策略包。

MVP：两周内做一个CLI/代理包装器：拦截shell和Git命令，用规则+LLM分类风险，设置审批预算与Slack/Webhook审批，对不可逆高危操作强制加锁，并输出审计日志。

竞品：LangChain Human-in-the-loop、OpenAI Agents SDK guardrails、Guardrails AI、Anthropic Claude Code permissions

摘要：作者公开DashClaw，一个让无人值守AI代理在执行破坏性操作前请求审批的层。日志显示7天产生1759次审批打断（约251次/天）却只有2次阻止，导致审批疲劳；调优后阈值又涨到101.8，超过0-100范围，模拟的drop production database cluster得95分仍被放行。作者因此加入不可被学习阈值覆盖的高风险下限，说明AI代理审批与风险分级存在真实产品机会。