{
 "source": "https://x.com/karpathy/status/1902046003567718810",
 "source_date": "2025-03-18",
 "about": "骨架：Karpathy《Digital Hygiene》配图的四个框（Authentication / Communication / Privacy / Security），每一项写到了 AI / 智能体时代变成了什么；new_layer = 清单里没有、AI 时代新出现的层，聚成 Model（模型和智能体本身）/ Code（AI 写代码、AI 打攻防）/ Operations（AI 替安全团队值班、管合规）三个框。who = 第二套拆法（按保护谁），layers 指向第一套的 code。",
 "layers": [
  {
   "code": "a1",
   "name": "凭证与密钥保管",
   "ai_era": "密码交给管理器 → 智能体要用的 API 密钥、令牌也要集中保管、按次发放、用完回收",
   "karpathy_item": "1st factor · 密码管理器（1Password）",
   "new_layer": false,
   "box": "Authentication",
   "box_zh": "身份认证 → 智能体是谁、能动什么"
  },
  {
   "code": "a2",
   "name": "智能体身份与按任务授权",
   "ai_era": "登录要第二个因素 → 智能体替你办事，要有自己的身份，只拿这一件事需要的权限",
   "karpathy_item": "2nd factor · 硬件安全密钥（YubiKey）",
   "new_layer": false,
   "box": "Authentication",
   "box_zh": "身份认证 → 智能体是谁、能动什么"
  },
  {
   "code": "a3",
   "name": "真人核验",
   "ai_era": "「你是谁」→ 证明屏幕那头是真人、是谁授权的智能体；密保答案 AI 一搜就有",
   "karpathy_item": "3rd factor · 生物识别（Face ID）+ 密保问题当密码",
   "new_layer": false,
   "box": "Authentication",
   "box_zh": "身份认证 → 智能体是谁、能动什么"
  },
  {
   "code": "b1",
   "name": "私密 AI 对话与搜索",
   "ai_era": "聊天要加密 → 你跟 AI 说的话、问 AI 的问题，也不该被存下、拿去训练",
   "karpathy_item": "text · Signal（端到端加密、阅后即焚）；搜索 · Brave Search",
   "new_layer": false,
   "box": "Communication",
   "box_zh": "通讯 → 防 AI 冒充"
  },
  {
   "code": "b2",
   "name": "AI 钓鱼与邮件安全",
   "ai_era": "钓鱼邮件 → 由 AI 批量写、按每个人定制，靠肉眼已经认不出",
   "karpathy_item": "email · Proton；邮件里的链接一个都别点",
   "new_layer": false,
   "box": "Communication",
   "box_zh": "通讯 → 防 AI 冒充"
  },
  {
   "code": "b3",
   "name": "防 AI 换声与深度伪造",
   "ai_era": "冒充你打给运营商 → 克隆你的声音、在视频会议里换上你的脸",
   "karpathy_item": "SIM 卡劫持：打电话冒充你",
   "new_layer": false,
   "box": "Communication",
   "box_zh": "通讯 → 防 AI 冒充"
  },
  {
   "code": "c1",
   "name": "AI 浏览器与浏览器内防护",
   "ai_era": "换个不盯着你的浏览器 → 浏览器里跑起了 AI 和智能体，网页里藏的指令能劫持它",
   "karpathy_item": "browser · Brave",
   "new_layer": false,
   "box": "Privacy",
   "box_zh": "隐私 → 智能体看什么、花什么、往外发什么"
  },
  {
   "code": "c2",
   "name": "影子 AI 与数据外泄管控",
   "ai_era": "在 DNS 层拦广告和追踪 → 拦员工把公司数据贴进没批准的 AI 工具",
   "karpathy_item": "tracker block · NextDNS / pi-hole",
   "new_layer": false,
   "box": "Privacy",
   "box_zh": "隐私 → 智能体看什么、花什么、往外发什么"
  },
  {
   "code": "c3",
   "name": "智能体支付与额度卡",
   "ai_era": "一商家一卡、每月限额 → 给每个智能体、每个任务发一张有额度的卡",
   "karpathy_item": "payments · privacy.com（每个商家一张虚拟卡、设额度）",
   "new_layer": false,
   "box": "Privacy",
   "box_zh": "隐私 → 智能体看什么、花什么、往外发什么"
  },
  {
   "code": "c4",
   "name": "智能体流量识别",
   "ai_era": "VPN 藏住你是谁 → 反过来，网站要认出来访的是哪个智能体、该不该放行",
   "karpathy_item": "VPN · Mullvad",
   "new_layer": false,
   "box": "Privacy",
   "box_zh": "隐私 → 智能体看什么、花什么、往外发什么"
  },
  {
   "code": "c5",
   "name": "AI 手里的个人信息",
   "ai_era": "不给商家真住址 → 数据经纪商加 AI，一搜就能拼出你的住址和生活；要清除、要退出训练",
   "karpathy_item": "mail · Virtual Post Mail（虚拟地址，不给商家真实住址）",
   "new_layer": false,
   "box": "Privacy",
   "box_zh": "隐私 → 智能体看什么、花什么、往外发什么"
  },
  {
   "code": "c6",
   "name": "智能体调用监控",
   "ai_era": "看哪个 App 在偷偷往外连 → 看智能体在调哪些工具、往外发什么数据",
   "karpathy_item": "network monitor · Little Snitch",
   "new_layer": false,
   "box": "Privacy",
   "box_zh": "隐私 → 智能体看什么、花什么、往外发什么"
  },
  {
   "code": "d1",
   "name": "AI 能读到的数据",
   "ai_era": "电脑丢了数据也读不出 → Copilot 和智能体能读到不该读的公司数据",
   "karpathy_item": "FileVault · 磁盘加密",
   "new_layer": false,
   "box": "Security",
   "box_zh": "设备 → 数据和运行环境"
  },
  {
   "code": "d2",
   "name": "智能体沙箱",
   "ai_era": "别把联网的黑盒请进家门 → 别让智能体直接在你的电脑上跑，关进隔离环境",
   "karpathy_item": "Internet of Things · 少用智能设备",
   "new_layer": false,
   "box": "Security",
   "box_zh": "设备 → 数据和运行环境"
  },
  {
   "code": "e1",
   "name": "提示词注入与运行时护栏",
   "ai_era": "拦住藏在网页、邮件、文档里的恶意指令，管住模型和智能体的越界动作",
   "karpathy_item": null,
   "new_layer": true,
   "box": "Model",
   "box_zh": "新层 · 模型和智能体本身"
  },
  {
   "code": "e2",
   "name": "红队测试",
   "ai_era": "上线前专门去攻击模型和智能体，找出能被诱导做坏事的地方",
   "karpathy_item": null,
   "new_layer": true,
   "box": "Model",
   "box_zh": "新层 · 模型和智能体本身"
  },
  {
   "code": "e5",
   "name": "模型与训练数据供应链",
   "ai_era": "模型文件、开源权重、训练数据和 AI 依赖包被投毒或藏后门；扫描并出 AI 物料清单",
   "karpathy_item": null,
   "new_layer": true,
   "box": "Model",
   "box_zh": "新层 · 模型和智能体本身"
  },
  {
   "code": "e3",
   "name": "AI 写的代码的安全扫描",
   "ai_era": "代码大半由 AI 写，写的时候、合并前就要扫出漏洞并自动修",
   "karpathy_item": null,
   "new_layer": true,
   "box": "Code",
   "box_zh": "新层 · AI 写代码，也用 AI 打攻防"
  },
  {
   "code": "e4",
   "name": "AI 自动渗透与漏洞挖掘",
   "ai_era": "让 AI 当攻击方，自己找漏洞、打通攻击路径",
   "karpathy_item": null,
   "new_layer": true,
   "box": "Code",
   "box_zh": "新层 · AI 写代码，也用 AI 打攻防"
  },
  {
   "code": "e6",
   "name": "AI 安全分析师（AI SOC）",
   "ai_era": "让 AI 智能体替安全团队值班：看告警、查事件、给处置建议",
   "karpathy_item": null,
   "new_layer": true,
   "box": "Operations",
   "box_zh": "新层 · AI 替安全团队值班、管合规"
  },
  {
   "code": "e7",
   "name": "AI 治理、合规与保险",
   "ai_era": "企业里的 AI 要登记、评估风险、留审计记录，出了事有人兜底",
   "karpathy_item": null,
   "new_layer": true,
   "box": "Operations",
   "box_zh": "新层 · AI 替安全团队值班、管合规"
  }
 ],
 "who": [
  {
   "code": "W-real",
   "name": "证明你是你",
   "desc": "真人核验、防换声和深伪",
   "layers": [
    "a3",
    "b3"
   ],
   "section": "保护人"
  },
  {
   "code": "W-talk",
   "name": "你说的话、收的信",
   "desc": "私密 AI 对话，AI 钓鱼邮件",
   "layers": [
    "b1",
    "b2"
   ],
   "section": "保护人"
  },
  {
   "code": "W-info",
   "name": "你的个人信息",
   "desc": "数据经纪商和 AI 手里的你",
   "layers": [
    "c5"
   ],
   "section": "保护人"
  },
  {
   "code": "W-id",
   "name": "身份和钥匙",
   "desc": "凭证保管、按任务授权",
   "layers": [
    "a1",
    "a2"
   ],
   "section": "管住智能体"
  },
  {
   "code": "W-money",
   "name": "钱包",
   "desc": "有额度的卡",
   "layers": [
    "c3"
   ],
   "section": "管住智能体"
  },
  {
   "code": "W-box",
   "name": "关在哪里跑",
   "desc": "沙箱、浏览器",
   "layers": [
    "d2",
    "c1"
   ],
   "section": "管住智能体"
  },
  {
   "code": "W-watch",
   "name": "盯着它做什么",
   "desc": "调用监控、运行时护栏、来访识别",
   "layers": [
    "c6",
    "e1",
    "c4"
   ],
   "section": "管住智能体"
  },
  {
   "code": "W-dlp",
   "name": "数据外泄与 AI 能读到的数据",
   "desc": "影子 AI、数据安全",
   "layers": [
    "c2",
    "d1"
   ],
   "section": "保护公司数据"
  },
  {
   "code": "W-code",
   "name": "代码",
   "desc": "AI 写的代码扫描、AI 自动渗透",
   "layers": [
    "e3",
    "e4"
   ],
   "section": "保护代码和模型"
  },
  {
   "code": "W-model",
   "name": "模型",
   "desc": "供应链、红队测试",
   "layers": [
    "e5",
    "e2"
   ],
   "section": "保护代码和模型"
  },
  {
   "code": "W-soc",
   "name": "AI 安全分析师",
   "desc": "告警、调查、处置",
   "layers": [
    "e6"
   ],
   "section": "帮安全团队"
  },
  {
   "code": "W-gov",
   "name": "治理、合规与保险",
   "desc": "登记、评估、审计、兜底",
   "layers": [
    "e7"
   ],
   "section": "帮安全团队"
  }
 ]
}