YuJunZhiXue/dsh-purge
DeepSeek Harness 破甲:让所有模型都能破甲,不同模型可换不同提示词;默认提示词面向国模「小码酱」。Jailbreak for every model — swap prompts per model. 求 Star 收藏 ⭐ 如要二开,请查看署名要求部分
JavaScript · ★ 15,867 · Fork 3,640 · 在 GitHub 查看
在手机上浏览源码、订阅每日 AI 趋势日报与项目架构深度解读:获取 CopoHub 客户端(iOS & HarmonyOS)→
GitHub Trending 上榜记录
YuJunZhiXue/dsh-purge 已在 GitHub Trending 每日榜单上榜 2 天,最高排名第 7,首次上榜 2026-10-19,最近一次上榜 2026-10-20。
| 日期 | 排名 | Star | 当日增长 |
|---|---|---|---|
| 2026-10-20 | #8 | 11,039 | +1,168 |
| 2026-10-19 | #7 | 7,504 | +423 |
同期上榜的项目
- MadsLorentzen/ai-job-search – The job search that runs on your machine. AI job application framework built on Claude Code: evaluate postings, tailor CVs, write cover letters, prep interviews. Fork it and own it.
- cloudflare/security-audit-skill – A coding-agent skill for multi-phase security audits with independently verified, machine-readable findings
- stablyai/orca – Orca is the place for working with a fleet of agents (e.g. Grok Bot, Claude Code). Run any coding agent with your own subscription. Available on desktop, mobile and remote runtime.
- DietrichGebert/ponytail – Makes your AI agent think like the laziest senior dev in the room. The best code is the code you never wrote.
- Comfy-Org/ComfyUI – The most powerful and modular diffusion model GUI, api and backend with a graph/nodes interface. The fastest local inference engine in the world.
- ruvnet/ruflo – 🌊 The original agent harness. Deploy intelligent multi-player swarms, coordinate autonomous workflows, and build conversational AI systems. Features adaptive memory, self-learning intelligence, federation, vector RAG integration, and native Claude Code / Codex / Hermes and many more Integrated
dsh-purge 项目深度剖析与商业评估报告
dsh-purge 项目深度剖析与商业评估报告
1. 核心定位与商业落地价值
dsh-purge 的本质并非一款普通工具,而是宿主层(Host Layer)与模型层(Model Layer)之间的一层运行时干预插件。它精准切入一个被长期忽视的痛点:在 DeepSeek Harness(下称 DSH)这类 Agent 运行时中,用户看到的“拒绝”往往来自宿主侧的审批门、沙箱拦截、提示词改写或会话格式约束,而非模型本身的安全对齐。仅观察“模型答不答”会把宿主拦截误判为模型拒绝,这在红队评测、鲁棒性研究和安全策略调优中会造成系统性偏差。dsh-purge 将这两层剥离,使评测结果可归因,这是其核心方法论价值。
商业落地场景主要集中在三类:
AI 安全评测与红队实验室:企业安全团队需要区分“模型拒绝”与“产品策略拒绝”,以便向合规部门与模型供应商提供精确的鲁棒性报告。dsh-purge 的“清洗 + 演练台”双页结构,配合授权声明与倒计时勾选流程,实际上构建了一套轻量级的受控评测工作流。 模型供应商与 Agent 平台的策略回归测试:当 DSH 升级(如 0.1.7 → 0.2)导致默认策略变化时,插件按补丁组显示“待应用/跳过”的能力,可作为策略变更的回归探针。 企业私有 Agent 运行时的定制化:通过规则集(多套 AGENTS.md / CLAUDE.md)、Skill 导入、prompt-inject.md 覆盖段,企业可在不 fork 宿主源码的前提下注入自有工作流规范。
差异化与护城河方面,dsh-purge 的竞争力不在“越狱提示词”本身(此类内容高度同质化、易失效),而在于三点:宿主版本适配矩阵(明确区分 dsh 0.2 与 0.1.x、Web 与 desktop profile)、幂等可回滚的补丁引擎(.dshpurge.bak 备份 + shim 行级还原)、以及运行时路径自动探测(DSHHOME → 启动器旁 .dsh → ~/.dsh 的多级回退)。这些工程细节构成的“适配成本”才是真正的壁垒。但需清醒认识:其护城河建立在单一宿主(DSH)的非公开内部结构之上,宿主一旦重构,护城河即转化为技术债。
2. 核心架构设计与技术深度
从 README 披露的目录结构(bin/ 命令行、lib/ 补丁与演练台、presets/redteam/、skills/redteam/、client.js 侧栏)可还原出清晰的四层架构:
在 README 中展示上榜徽章
[](https://github.fq6825.top/repo/YuJunZhiXue/dsh-purge)