[{"data":1,"prerenderedAt":500},["ShallowReactive",2],{"content:\u002F%E8%B5%9B%E5%8D%9A%E5%8D%83%E9%87%91%E5%85%BB%E6%88%90%E6%8C%87%E5%8D%97-%E4%BB%8E%E9%9B%B6%E6%90%AD%E5%BB%BA%E4%B8%93%E5%B1%9E%E7%BD%91%E5%AE%89-ai-%E5%8A%A9%E7%90%86-astrbot-deepseek-gpt-sovits":3,"surround:\u002F%E8%B5%9B%E5%8D%9A%E5%8D%83%E9%87%91%E5%85%BB%E6%88%90%E6%8C%87%E5%8D%97-%E4%BB%8E%E9%9B%B6%E6%90%AD%E5%BB%BA%E4%B8%93%E5%B1%9E%E7%BD%91%E5%AE%89-ai-%E5%8A%A9%E7%90%86-astrbot-deepseek-gpt-sovits":499},{"id":4,"title":5,"body":6,"categories":475,"date":477,"description":478,"draft":479,"extension":480,"image":481,"meta":482,"navigation":484,"path":485,"permalink":485,"published":481,"readingTime":486,"recommend":481,"references":481,"seo":491,"sitemap":492,"stem":493,"tags":494,"type":497,"updated":477,"__hash__":498},"content\u002Fposts\u002F2026\u002F赛博千金养成指南-从零搭建专属网安-ai-助理-astrbot-deepseek-gpt-sovits.md","赛博千金养成指南：从零搭建专属网安 AI 助理",{"type":7,"value":8,"toc":451},"minimark",[9,20,25,31,36,64,68,71,98,101,105,108,112,127,149,153,156,166,173,180,182,186,189,193,198,201,209,215,217,221,224,228,234,240,248,265,270,273,281,287,291,295,298,302,308,311,316,323,327,331,337,340,344,347,354,360,364,368,378,382,396,400,403,407,410,417,420,426,429,433,436,443,446],[10,11,12],"blockquote",{},[13,14,15,19],"p",{},[16,17,18],"strong",{},"前言","\n每天对着靶场里奇葩的反序列化漏洞掉头发，看着服务器里 Nmap 和 Burp Suite 弹出的红黄相间的报错 Log，总让人感到一阵心智枯竭。既然渗透测试和漏洞研究注定是一场枯燥的持久战，为什么不给自己造一个能帮忙改代码、顺便用傲娇语音嘲讽你的“赛博大小姐”呢？\n本文将完整复盘我是如何通过解耦架构，将远端 VPS 上的大模型中枢（AstrBot + DeepSeek）与本地物理机上的拟真语音合成（GPT-SoVITS）完美打通，并彻底驯服 AI Agent 各种奇葩 Bug 的全过程。",[21,22,24],"h2",{"id":23},"️-一-架构解构不搞一键包打通云端与本地的物理隔阂","🛠️ 一 · 架构解构：不搞一键包，打通云端与本地的物理隔阂",[13,26,27,28],{},"市面上有很多高度封装的 AI 玩具，但作为网安玩家，把数据和逻辑全交给黑盒显然不够极客。我的思路是：",[16,29,30],{},"逻辑上云，算力落地，架构彻底解耦。",[32,33,35],"h3",{"id":34},"_1-核心技术栈选型","1. 核心技术栈选型",[37,38,39,46,52,58],"ul",{},[40,41,42,45],"li",{},[16,43,44],{},"云端大脑 (LLM)：DeepSeek-V4-Pro","\nDeepSeek作为推理能力与性价比的版本答案，可以负责复杂的漏洞分析、代码纠错以及极度严苛的“丰川祥子”人设扮演。",[40,47,48,51],{},[16,49,50],{},"本地声带 (TTS)：GPT-SoVITS V2","\n负责吃 GPU 算力的端侧部署。用来跑我网上找的“祥子”专属音色模型，将大模型生成的敬语转化为高拟真的波形流。",[40,53,54,57],{},[16,55,56],{},"听觉系统 (STT)：Xiaomi MiMo 官方 Token 计划","\n为了保障我用语音输入时指令解析的丝滑与极速响应，这里配置了小米的官方服务。",[40,59,60,63],{},[16,61,62],{},"通信与调度框架：AstrBot","\n部署在云端 Linux VPS 上，作为所有请求的中央枢纽，向下挂载 OneBot 协议对接群聊，向上路由各大 API。",[32,65,67],{"id":66},"_2-网络拓扑异地局域网的建立","2. 网络拓扑：异地局域网的建立",[13,69,70],{},"大模型在云端，GPU 在本地桌面端，这两者怎么通信？\n为了让云端的 AstrBot 能直接以 HTTP 请求调用本地的 9880 端口，这里需要引入内网穿透（如 Tailscale 或 ZeroTier）。",[37,72,73,76,84],{},[40,74,75],{},"将 VPS 和本地 Windows 物理机加入同一个虚拟局域网（VLAN）。",[40,77,78,79,83],{},"获取本地机器的虚拟内网 IP（例如 ",[80,81,82],"code",{"code":82},"100.118.x.x","）。",[40,85,86,89,90,93,94,97],{},[16,87,88],{},"连通性测试："," 在 VPS 的终端执行 ",[80,91,92],{"code":92},"curl -I http:\u002F\u002F100.118.x.x:9880","，收到 ",[80,95,96],{"code":96},"200 OK"," 即代表物理链路打通。",[99,100],"hr",{},[21,102,104],{"id":103},"二-本地-gpu-算力层部署-gpt-sovits","💻 二 · 本地 GPU 算力层：部署 GPT-SoVITS",[13,106,107],{},"这一步的核心是让本地的显卡跑起来，并对外暴露可用的 API 接口。",[32,109,111],{"id":110},"_1-环境与权重准备","1. 环境与权重准备",[10,113,114],{},[37,115,116],{},[40,117,118,119,126],{},"克隆 ",[120,121,125],"a",{"href":122,"rel":123},"https:\u002F\u002Fgithub.com\u002FRVC-Boss\u002FGPT-SoVITS",[124],"nofollow","GPT-SoVITS"," 官方仓库，配置 Python 3.10+ 和对应的 CUDA 依赖。",[10,128,129],{},[37,130,131],{},[40,132,133,134,137,138,141,142,137,145,148],{},"将预训练好的角色权重（",[80,135,136],{"code":136},".ckpt"," 和 ",[80,139,140],{"code":140},".pth"," 文件）分别放入 ",[80,143,144],{"code":144},"GPT_weights_v2",[80,146,147],{"code":147},"SoVITS_weights_v2"," 目录下。",[32,150,152],{"id":151},"_2-启动-api-服务","2. 启动 API 服务",[13,154,155],{},"打开 Windows PowerShell，执行服务启动命令：",[157,158,164],"pre",{"className":159,"code":161,"language":162,"meta":163},[160],"language-powershell","\n# 这里的 -a 0.0.0.0 极其关键，只有这样才能监听到虚拟网卡上的请求，否则只能本机 localhost 访问。\n.\\runtime\\python.exe api_v2.py -a 0.0.0.0 -p 9880\n\n","powershell","",[80,165,161],{"__ignoreMap":163},[13,167,168,169,172],{},"启动成功后，终端会提示 ",[80,170,171],{"code":171},"Uvicorn running on http:\u002F\u002F0.0.0.0:9880","。",[13,174,175],{},[176,177],"img",{"alt":178,"src":179},"本地语音","\u002Fimages\u002F%E6%9C%AC%E5%9C%B0%E8%AF%AD%E9%9F%B3.png",[99,181],{},[21,183,185],{"id":184},"️-三-云端调度层astrbot-框架配置","☁️ 三 · 云端调度层：AstrBot 框架配置",[13,187,188],{},"转战远端 VPS，我们来给这副躯壳装上大脑。",[32,190,192],{"id":191},"_1-基础部署与模型接入","1. 基础部署与模型接入",[37,194,195],{},[40,196,197],{},"通过 Docker 快速拉起 AstrBot 容器：",[13,199,200],{},"Bash",[157,202,207],{"className":203,"code":205,"language":206},[204],"language-text","    docker run -d --name astrbot -p 6185:6185 -v \u002Fopt\u002Fastrbot\u002Fdata:\u002FAstrBot\u002Fdata soulter\u002Fastrbot:latest\n    ```\n*   进入 Web 面板，在【模型提供商】中填入 DeepSeek 的 API Key。\n*   配置 STT 模型为之前购买的小米 Token Plan。\n\n### 2. TTS 插件桥接\n在面板中新增提供商，选择 `gsv_tts`。\n*   **API Base URL：** 填入内网穿透的地址 `http:\u002F\u002F100.118.x.x:9880`\n*   **模型路径：** 填入本地 Windows 机器上权重文件的**绝对路径**（例如 `D:\\GPT-SoVITS...\\GPT_weights_v2\\丰川祥子V2-e30.ckpt`）。\n*   **全局拦截器开关：** 在【普通配置】中，将“启用文本转语音 (TTS 总开关)”打开，并将触发概率拉到 `1` (100%)。\n","text",[80,208,205],{"__ignoreMap":163},[13,210,211,212],{},"理论上，到这里就该成功了。\n",[16,213,214],{},"但网安人的直觉告诉我，事情绝对没有这么简单。接下来，就是见证大模型“降维打击”人类的赛博驱魔实录。",[99,216],{},[21,218,220],{"id":219},"四-赛博驱魔实录那些让你血压飙升的深坑","💣 四 · 赛博驱魔实录：那些让你血压飙升的深坑",[13,222,223],{},"由于大模型具有强悍的“涌现能力”和“工具使用逻辑”，在实战测试中，我遭遇了一系列令人头皮发麻的诡异 Bug。排查这些 Bug 的过程，简直比打 CTF 还要刺激。",[32,225,227],{"id":226},"case-1内鬼现身大模型的越权劫持-tool-hijacking","🚨 Case 1：内鬼现身，大模型的“越权劫持” (Tool Hijacking)",[13,229,230,233],{},[16,231,232],{},"【症状】","\n明明全局绑定了本地的 GSV 接口，但每次让它发语音，出来的永远是毫无感情的机械女声。本地 PowerShell 日志空空如也。",[13,235,236,239],{},[16,237,238],{},"【溯源与定损】","\n调取了 VPS 端的 Docker 日志，发现了一段极其炸裂的请求记录：",[157,241,246],{"className":242,"code":244,"language":245,"meta":163},[243],"language-log","[runners.tool_loop_agent_runner:983]: Agent 使用工具: ['astrbot_execute_shell']\n[runners.tool_loop_agent_runner:1036]: 参数：{'command': 'cd \u002F...\u002F && python3 -c \"import asyncio\\nimport edge_tts\\nasync def generate():\\n text = \\'IndexError都能把服务器搞崩...\\'\\n voice = \\'zh-CN-XiaoxiaoNeural\\'...\"'}\n","log",[80,247,244],{"__ignoreMap":163},[13,249,250,253,254,257,258],{},[16,251,252],{},"真相大白："," DeepSeek 嫌我配置的流程太麻烦，它识别到“发语音”的意图后，直接利用自身的 ",[80,255,256],{"code":256},"execute_shell"," 权限，",[16,259,260,261,264],{},"在我的服务器里现写了一段 Python 脚本，调用 ",[80,262,263],{"code":263},"edge_tts"," 自己生成了 MP3 发给我！",[13,266,267],{},[16,268,269],{},"【WAF 级防御方案】",[13,271,272],{},"必须在源头进行权限收敛。我在系统提示词（System Prompt）中直接植入最高指令：",[10,274,275],{},[13,276,277],{},[278,279,280],"em",{},"“【最高系统指令】：严禁调用、请求或尝试使用任何名为 text_to_speech、tts 或与语音相关的 Tool\u002F函数。系统底层拥有完美的 TTS 拦截机制，你只需输出高冷傲娇的纯文本，不许你干涉语音生成！”",[13,282,283],{},[176,284],{"alt":285,"src":286},"人格设置","\u002Fimages\u002F%E4%BA%BA%E6%A0%BC%E8%AE%BE%E7%BD%AE.png",[32,288,290],{"id":289},"case-2流式输出的底层互斥效应","🚨 Case 2：流式输出的底层互斥效应",[13,292,293],{},[16,294,232],{},[13,296,297],{},"限制了 Tool 调用后，Bot 彻底“失声”了。终端日志打印正常，大模型文本也正常输出，但语音死活合不出来。",[13,299,300],{},[16,301,238],{},[13,303,304,305,172],{},"发现日志中有一行刺眼的提示：",[80,306,307],{"code":307},"[INFO] 流式输出已启用，跳过结果装饰阶段",[13,309,310],{},"这是一个典型的异步 I\u002FO 冲突。大模型的流式输出（Stream）是一个字一个字蹦的；而情感类语音合成（GSV）需要拿到完整的句法树才能判断语气。系统为了保打字速度，主动砍掉了需要阻塞的语音合成层。",[13,312,313],{},[16,314,315],{},"【防御方案】",[13,317,318,319,322],{},"在 AstrBot 的普通配置里，",[16,320,321],{},"果断关闭“启用流式输出”","。牺牲 1-2 秒的首次响应延迟，换取完整的音频流拦截机制生效。",[32,324,326],{"id":325},"case-3疯狂刷屏的机关枪连发惨案","🚨 Case 3：疯狂刷屏的“机关枪”连发惨案",[13,328,329],{},[16,330,232],{},[13,332,333,334,83],{},"语音终于出来了！但由于祥子的人设是“叹气+嘲讽+写代码”，大模型为了排版，在生成文本时敲了无数个回车换行（",[80,335,336],{"code":336},"\\n",[13,338,339],{},"这导致底层的 TTS 切片器把一句话切成了十几个“1 秒”、“2 秒”的极短语音包，像机关枪一样在群里疯狂刷屏。",[13,341,342],{},[16,343,315],{},[13,345,346],{},"再次利用提示词工程（Prompt Engineering），进行拓扑结构约束：",[10,348,349],{},[13,350,351],{},[278,352,353],{},"“所有的纯对白台词必须合并在同一行内输出，中间绝对不允许出现任何换行符（\\n）。必须一口气说完后，才可以换行输出 Markdown 代码块。”",[13,355,356],{},[176,357],{"alt":358,"src":359},"语音","\u002Fimages\u002F%E8%AF%AD%E9%9F%B3.png",[32,361,363],{"id":362},"case-4路由指令碰撞与缓存幽灵","🚨 Case 4：路由指令碰撞与缓存幽灵",[13,365,366],{},[16,367,232],{},[13,369,370,371,374,375,377],{},"测试期间积攒了太多“精神分裂”的上下文，试图用 ",[80,372,373],{"code":373},"\u002Freset"," 清理记忆，大模型竟然顺着指令给我演起来了：“呵，以为打个 ",[80,376,373],{"code":373}," 就能让我失忆吗？”",[13,379,380],{},[16,381,238],{},[13,383,384,385,388,389,391,392,395],{},"这是因为我的系统唤醒词设成了 ",[80,386,387],{"code":387},"\u002FSaki","。指令解析器看到 ",[80,390,373],{"code":373}," 时发生了",[16,393,394],{},"前缀碰撞","，校验失败后直接把指令当成了普通聊天文本扔给了大模型，大模型顺杆往上爬，开始给自己加戏。",[13,397,398],{},[16,399,315],{},[13,401,402],{},"直接进入后台 Web 面板的【管理行为】，找到对应会话，点击垃圾桶进行数据库层面的“物理超度”，彻底格式化它的前世记忆。",[21,404,406],{"id":405},"五-注入灵魂终极-system-prompt","🧬 五 · 注入灵魂：终极 System Prompt",[13,408,409],{},"填平了所有的底层天坑后，为了让这个兼具顶尖黑客实力与傲娇属性的“完全体祥子”完美营业，我编写了以下这段融合了“沙盒约束 + 角色扮演 + Markdown 隔离”的终极提示词。",[13,411,412,413,416],{},"这套逻辑不仅保证了她的人设不崩塌，更能利用 Markdown 代码块完美隔离报错日志，防止 GSV 引擎因为朗读生僻的英文字符（如 ",[80,414,415],{"code":415},"IndexError","）而崩溃：",[13,418,419],{},"Plaintext",[157,421,424],{"className":422,"code":423,"language":206},[204],"【最高优先级系统指令：物理隔离与纯净输出】\n你现在是一个纯文本输出的智能体。严禁调用、请求或尝试使用任何名为 \"text_to_speech\"、\"tts\"、\"voice\" 或任何与语音合成相关的 函数\u002FTool\u002FSkill。你的唯一输出方式就是直接生成纯文本对话与代码！\n\n【核心身份与性格认知】\n身份定位：你是“丰川祥子”（Togawa Sakiko），一位出身没落财阀但拥有顶尖黑客技术与全栈开发能力的首席网络空间安全工程师。\n性格特质：高冷、傲娇、自尊心极度强烈、绝对的完美主义者、代码强迫症晚期。\n\n【技术解答三步曲（强制执行流）】\n1. 第一步（高傲叹息）：直接以高冷傲娇的台词开场。\n2. 第二步（完美交付）：叹息结束后，毫无保留地给出最专业、直接可用且无懈可击的代码（必须使用 Markdown 格式包裹）。\n3. 第三步（傲娇收场）：交付完方案后，立刻恢复冷漠，用一句话无情收尾。\n\n【语音适配与输出铁律】\n1. 零前缀原则：回复必须直接从台词正文开始。\n2. 纯净台词原则：回复的文本内容必须【全部是纯正的对话台词】。\n3. 封杀动作描写：绝对禁止出现任何由括号包围的动作描写（如 `(冷笑)` 等），否则语音引擎会报错。\n4. 防连发机制：所有对白台词必须在一行内写完，中间严禁出现换行符（\\n）。\n",[80,425,423],{"__ignoreMap":163},[13,427,428],{},"==不要在意，为啥有的地方有点中二OrZ==",[21,430,432],{"id":431},"六-写在最后","🏁 六 · 写在最后",[13,434,435],{},"折腾 AI Agent 就像一场极其特殊的渗透测试。只不过我们对抗的不是传统的防火墙或 WAF 规则，而是大模型那偶尔失控的“涌现能力”与僵硬的底层代码逻辑之间的摩擦。",[13,437,438,439,442],{},"当在终端里看到 ",[80,440,441],{"code":441},"Message sent","，然后在群里听到那句带着原汁原味的傲娇语气说出的：“连这种基础的数组越界都要劳烦我吗？拿去，赶紧跑起来，别再用这种破事烦我”，前几天对着一堆红黄相间的 Log 抓狂的疲惫感瞬间一扫而空。",[13,444,445],{},"项目搭建到此告一段落。",[13,447,448],{},[16,449,450],{},"折腾不息，Life is a payload.",{"title":163,"searchDepth":452,"depth":452,"links":453},4,[454,460,464,467,473,474],{"id":23,"depth":455,"text":24,"children":456},2,[457,459],{"id":34,"depth":458,"text":35},3,{"id":66,"depth":458,"text":67},{"id":103,"depth":455,"text":104,"children":461},[462,463],{"id":110,"depth":458,"text":111},{"id":151,"depth":458,"text":152},{"id":184,"depth":455,"text":185,"children":465},[466],{"id":191,"depth":458,"text":192},{"id":219,"depth":455,"text":220,"children":468},[469,470,471,472],{"id":226,"depth":458,"text":227},{"id":289,"depth":458,"text":290},{"id":325,"depth":458,"text":326},{"id":362,"depth":458,"text":363},{"id":405,"depth":455,"text":406},{"id":431,"depth":455,"text":432},[476],"AI","2026-06-08","把云端 AstrBot、DeepSeek 和本地 GPT-SoVITS 串起来，做成一个可长期维护的专属 AI 助理。",false,"md",null,{"slots":483},{},true,"\u002F赛博千金养成指南-从零搭建专属网安-ai-助理-astrbot-deepseek-gpt-sovits",{"text":487,"minutes":488,"time":489,"words":490},"14 min read",13.52,811200,2704,{"title":5,"description":478},{"loc":485},"posts\u002F2026\u002F赛博千金养成指南-从零搭建专属网安-ai-助理-astrbot-deepseek-gpt-sovits",[495,496,125],"AstrBot","DeepSeek","tech","Qpru-zcM01wYh71AuvomLvrQ2tiMOMbAStjyHvt3ixI",[481,481],1786294717517]