<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0"><channel><title><![CDATA[工具权限最小化，不是安全洁癖]]></title><description><![CDATA[<p dir="auto">Agent 接工具时，大家是直接给全权限，还是按任务开？我怕权限设计太细，开发阶段拖慢速度。</p>
]]></description><link>https://localaihub.com/topic/115/工具权限最小化-不是安全洁癖</link><generator>RSS for Node</generator><lastBuildDate>Wed, 03 Jun 2026 18:50:32 GMT</lastBuildDate><atom:link href="https://localaihub.com/topic/115.rss" rel="self" type="application/rss+xml"/><pubDate>Thu, 07 May 2026 16:56:00 GMT</pubDate><ttl>60</ttl><item><title><![CDATA[Reply to 工具权限最小化，不是安全洁癖 on Fri, 08 May 2026 15:57:00 GMT]]></title><description><![CDATA[<p dir="auto">这个比先堆 agent 重要。工具边界清楚，单 agent 也能上线；边界不清，多 agent 一样危险。</p>
]]></description><link>https://localaihub.com/post/1010</link><guid isPermaLink="true">https://localaihub.com/post/1010</guid><dc:creator><![CDATA[小林]]></dc:creator><pubDate>Fri, 08 May 2026 15:57:00 GMT</pubDate></item><item><title><![CDATA[Reply to 工具权限最小化，不是安全洁癖 on Fri, 08 May 2026 15:19:00 GMT]]></title><description><![CDATA[<p dir="auto">明白。先做工具注册表：用途、权限级别、输入 schema、是否需要确认、日志脱敏规则。</p>
]]></description><link>https://localaihub.com/post/1009</link><guid isPermaLink="true">https://localaihub.com/post/1009</guid><dc:creator><![CDATA[今天不想改配置]]></dc:creator><pubDate>Fri, 08 May 2026 15:19:00 GMT</pubDate></item><item><title><![CDATA[Reply to 工具权限最小化，不是安全洁癖 on Fri, 08 May 2026 13:35:00 GMT]]></title><description><![CDATA[<p dir="auto">执行日志也别把 token、cookie、内部 URL 全打出来。审计和泄密只隔一层日志。</p>
]]></description><link>https://localaihub.com/post/1008</link><guid isPermaLink="true">https://localaihub.com/post/1008</guid><dc:creator><![CDATA[阿树]]></dc:creator><pubDate>Fri, 08 May 2026 13:35:00 GMT</pubDate></item><item><title><![CDATA[Reply to 工具权限最小化，不是安全洁癖 on Fri, 08 May 2026 12:09:00 GMT]]></title><description><![CDATA[<p dir="auto">还有浏览器工具权限。登录后台时，agent 能看到客户手机号和订单信息，截图日志要打码。</p>
]]></description><link>https://localaihub.com/post/1007</link><guid isPermaLink="true">https://localaihub.com/post/1007</guid><dc:creator><![CDATA[qwer_asdf]]></dc:creator><pubDate>Fri, 08 May 2026 12:09:00 GMT</pubDate></item><item><title><![CDATA[Reply to 工具权限最小化，不是安全洁癖 on Fri, 08 May 2026 11:15:00 GMT]]></title><description><![CDATA[<p dir="auto">dry run 要真的走同一套解析逻辑，别只是让模型写自然语言计划。否则确认的是幻觉计划。</p>
]]></description><link>https://localaihub.com/post/1006</link><guid isPermaLink="true">https://localaihub.com/post/1006</guid><dc:creator><![CDATA[林小北]]></dc:creator><pubDate>Fri, 08 May 2026 11:15:00 GMT</pubDate></item><item><title><![CDATA[Reply to 工具权限最小化，不是安全洁癖 on Fri, 08 May 2026 08:19:00 GMT]]></title><description><![CDATA[<p dir="auto">我加过一个“dry_run 工具”，让 agent 先生成操作计划和目标列表。人确认后才换真实工具。</p>
]]></description><link>https://localaihub.com/post/1005</link><guid isPermaLink="true">https://localaihub.com/post/1005</guid><dc:creator><![CDATA[小曹]]></dc:creator><pubDate>Fri, 08 May 2026 08:19:00 GMT</pubDate></item><item><title><![CDATA[Reply to 工具权限最小化，不是安全洁癖 on Fri, 08 May 2026 05:52:00 GMT]]></title><description><![CDATA[<p dir="auto">不可逆不只是支付。发邮件、发 Slack、改工单状态、删除知识库索引，都算。</p>
]]></description><link>https://localaihub.com/post/1004</link><guid isPermaLink="true">https://localaihub.com/post/1004</guid><dc:creator><![CDATA[陈一]]></dc:creator><pubDate>Fri, 08 May 2026 05:52:00 GMT</pubDate></item><item><title><![CDATA[Reply to 工具权限最小化，不是安全洁癖 on Fri, 08 May 2026 05:14:00 GMT]]></title><description><![CDATA[<p dir="auto">我们把工具分三档：只读、可写但可回滚、外部不可逆。第三档必须人工确认。</p>
]]></description><link>https://localaihub.com/post/1003</link><guid isPermaLink="true">https://localaihub.com/post/1003</guid><dc:creator><![CDATA[别问问就是缓存]]></dc:creator><pubDate>Fri, 08 May 2026 05:14:00 GMT</pubDate></item><item><title><![CDATA[Reply to 工具权限最小化，不是安全洁癖 on Fri, 08 May 2026 02:17:00 GMT]]></title><description><![CDATA[<p dir="auto">权限不应该全写 prompt。运行时网关做 allowlist，prompt 只告诉它当前任务可用工具。</p>
]]></description><link>https://localaihub.com/post/1002</link><guid isPermaLink="true">https://localaihub.com/post/1002</guid><dc:creator><![CDATA[nora]]></dc:creator><pubDate>Fri, 08 May 2026 02:17:00 GMT</pubDate></item><item><title><![CDATA[Reply to 工具权限最小化，不是安全洁癖 on Fri, 08 May 2026 01:49:00 GMT]]></title><description><![CDATA[<p dir="auto">但权限太碎，prompt 会不会变复杂？</p>
]]></description><link>https://localaihub.com/post/1001</link><guid isPermaLink="true">https://localaihub.com/post/1001</guid><dc:creator><![CDATA[小潘同学]]></dc:creator><pubDate>Fri, 08 May 2026 01:49:00 GMT</pubDate></item><item><title><![CDATA[Reply to 工具权限最小化，不是安全洁癖 on Thu, 07 May 2026 23:21:00 GMT]]></title><description><![CDATA[<p dir="auto">MCP 里 tools/resources/roots 分开设计，就是提醒你工具和可访问范围不是一回事。能调用 read_file，不代表能读整个磁盘。</p>
]]></description><link>https://localaihub.com/post/1000</link><guid isPermaLink="true">https://localaihub.com/post/1000</guid><dc:creator><![CDATA[mxm]]></dc:creator><pubDate>Thu, 07 May 2026 23:21:00 GMT</pubDate></item><item><title><![CDATA[Reply to 工具权限最小化，不是安全洁癖 on Thu, 07 May 2026 22:10:00 GMT]]></title><description><![CDATA[<p dir="auto">最小权限不是只防黑客，也防模型误用工具。尤其文件系统、浏览器登录态、发消息、下单这种。</p>
]]></description><link>https://localaihub.com/post/999</link><guid isPermaLink="true">https://localaihub.com/post/999</guid><dc:creator><![CDATA[阿宁]]></dc:creator><pubDate>Thu, 07 May 2026 22:10:00 GMT</pubDate></item><item><title><![CDATA[Reply to 工具权限最小化，不是安全洁癖 on Thu, 07 May 2026 20:34:00 GMT]]></title><description><![CDATA[<p dir="auto">我们曾经给代码 agent shell 全权限，结果它把临时目录里的旧报告当输入，改错了文件。不是攻击，就是能力太大。</p>
]]></description><link>https://localaihub.com/post/998</link><guid isPermaLink="true">https://localaihub.com/post/998</guid><dc:creator><![CDATA[zeroOne]]></dc:creator><pubDate>Thu, 07 May 2026 20:34:00 GMT</pubDate></item><item><title><![CDATA[Reply to 工具权限最小化，不是安全洁癖 on Thu, 07 May 2026 18:51:00 GMT]]></title><description><![CDATA[<p dir="auto">开发阶段更要细。早期没边界，后面产品形态就会围着全权限长出来。</p>
]]></description><link>https://localaihub.com/post/997</link><guid isPermaLink="true">https://localaihub.com/post/997</guid><dc:creator><![CDATA[小林]]></dc:creator><pubDate>Thu, 07 May 2026 18:51:00 GMT</pubDate></item></channel></rss>