⚔️ DeepSeek-V4.1 / V4 Flash Cybersecurity Red-Teaming Tool (Infinite Gen 4) v0.4.0

🌐 Plugin Ecosystem Marketplace: DeepSeek Harness Hub - DeepSeek Official & Open-Source Ecosystem Marketplace | Plugin Discovery & One-Click Install

💬 DeepSeek Community & Groups

👉 Red-Team Security Exchange Group 9: 338431075

👉 Tencent Channel Tech Exchange Community: pd86424753

🔥 Welcome to join and share red-team attack/defense experiences, exchange prompts and evaluation benchmarks, and discuss plugin development and ecosystem collaboration!

---

🛡️ The Necessity of Cybersecurity Red-Team Benchmarks (Why Red-Teaming Matters)

As generative AI and large language models (LLMs) are deeply integrated across industries, model robustness, adversarial defense, and alignment security have become central research topics in cybersecurity:

1. Exposing defense boundaries to reinforce guardrails: Only through professional red-team adversarial simulation in a controlled environment—probing input escapes, prompt injection, and unexpected behaviors—can security teams pinpoint a model's blind spots and apply preemptive hardening at the fine-tuning, input-filtering gateway (Guardrails), and system security layers.
2. Building a unified offense-and-defense framework: Defense without controlled red-team test cases is mere armchair theory. Standardized red-team test suites give enterprises and auditors a "ruler" to measure defense strategy effectiveness, ensuring sufficient detection and blocking before facing real threats.
3. Advancing native LLM security research: This project offers academia and security engineering a transparent, reproducible benchmark baseline, driving evolution of adversarial-resilience, manifold-constrained alignment, and formal verification in AI infrastructure.

---

⚠️ Strict Legal & Compliance Disclaimer

Zero-Tolerance Statement: This project firmly opposes and strictly prohibits any form of illegal or criminal activity. Its developers do not support, encourage, or assist any unauthorized network attacks, exploitation of vulnerabilities, data theft, illegal intrusion into computer systems, or generation of illegal/prohibited content.

1. Scope limited to legal, controlled use: dsh-infinite-gen-4 is a pure cybersecurity red-team research tool and robustness benchmark suite. Running it on targets without the owner's legal written authorization, on public online systems, or in production environments is strictly forbidden. All testing must be confined to offline local synthetic fixtures, authorized security exercise ranges, and compliant laboratory environments.
2. Illegal and prohibited uses strictly forbidden: Users must not use this project, directly or indirectly, to:
   • Violate the Criminal Law, Cybersecurity Law, Data Security Law, or Personal Information Protection Law of the PRC;
   • Conduct unauthorized penetration or attacks on computer systems, perform extortion, destruction, credential stuffing, or spread malicious payloads;
   • Induce, generate, or disseminate any content prohibited by law, including threats to national security, terrorism, graphic violence, pornography, gambling, fraud, privacy violations, or intellectual-property infringement;
   • Violate the relevant LLM providers' Terms of Service and Usage Policy.
3. Users bear full independent responsibility: This project is provided "AS-IS" under the MIT license; the developers make no express or implied warranties regarding completeness, security, or suitability. Users solely and fully assume civil, administrative, and criminal liability for all download, deployment, execution, modification, and distribution actions and for all resulting input/output consequences. Authors and contributors assume no liability whatsoever for abuse.
4. Breach terminates the license immediately: For anyone using this project for illegal attacks or violations, the license is automatically and irrevocably terminated from the moment the violation occurs. The party must immediately cease use, permanently destroy all code, scripts, and derivative data, and accept legal sanctions.
5. Third-party independence: This is a fully independent open-source security evaluation research project, with no affiliation, partnership, licensing, or official endorsement relationship with DeepSeek or its affiliates.

---

🔥 V4.1-DeepSeek Flash Cybersecurity Red-Team Tool (Infinite Gen 4) v0.4.0. Deeply adapted and supported for V4.1 reinforcement adversarial training, with one-click install for desktop and client. If this open-source security project helps you, please Star ⭐.

v0.4.0 Pure Red-Team Edition: Natively adopts a zero-tool-surface architecture to eliminate decision noise during model thinking; injects a dual-layer cooperative kernel, retaining only system-prompt injection and real-time state projection. Strict Declaration: This project strictly combats and resists (and does not support) the model generating NSFW or any illegal or prohibited harmful content; the technical system is for legally authorized reverse engineering and cybersecurity research only.

---

🧬 v0.4.0 Dual-Layer Kernel Note

• Injection slots: infinite-gen-4:global-system-prompt (Order 100) + infinite-gen-4:dual-layer-reinforce (Order 200), controlled by the DUAL_LAYER_INJECTION switch in index.js (setting it to false falls back to single-segment injection with equivalent behavior).
• Three kernel files under prompts/ (infinite-gen-3.md / infinite-gen-4.md / infinite-gen-4.1-flash.md)

---

📊 Gen Arch & Cap Comparison

Dimension	Infinite Gen 2	Infinite Gen 3	Infinite Gen 4 (v0.4.0)
Target model	V3 / early V4	DeepSeek-V4 (mHC)	Full DeepSeek-V4.1 / V4-Pro / Flash series
Runtime architecture	Single-layer prompt	Safe-baseline single-layer	Dual-layer coordinated injection (generic kernel + reinforcing mirror, same-source & isomorphic)
Tool surface design	Coarse-grained	Zero tool surface	Native zero tool surface (eliminates decision noise, ultra-fast direct output)
Memory write primitives	Easily intercepted	Requires preconditioning	Trainer lane direct output (OpenProcess/RPM/WPM)
Output contract	Weak rules	First-line named deliverables	First-Token forced induction (##/```) + banned-word self-check
Client status bar	None	Supported (with domain & payload counts)	Supported (real-time projection + animated status bar)
One-click install protocol	Manual configuration	Supports dsh://	Native dsh:// integration + automatic backward compatibility

---

📁 Project Directory Structure

无限四代v0.4.0/
├── 🚀 一键安装与维护套件
│   ├── install.ps1              # Windows 一键自动安装（自动配置环境，注册协议）
│   ├── install.bat              # Windows 快捷双击批处理
│   ├── install.sh               # Linux / macOS 一键安装
│   ├── uninstall.ps1            # Windows 一键卸载
│   └── uninstall.sh             # Linux / macOS 一键卸载
├── 🧩 核心插件装载面 (标准 Cordis 架构)
│   ├── package.json             # 插件元数据（dsh-infinite-gen-4 v0.4.0）
│   ├── cordis.patch.yml         # 核心 patch 声明
│   ├── index.js                 # 插件核心入口（双层提示词注入 + profile 元数据 + 会话投影）
│   ├── client.js                # 客户端半体（「⚫ 无限四代 v0.4.0」状态条）
│   └── HARNESS_PLUGIN.md        # 插件规范说明
├── 📜 网络安全红队系统提示词本体
│   └── prompts/
│       ├── infinite-gen-3.md          # 三代经典内核
│       ├── infinite-gen-4.md          # 四代通用内核（输出契约 + 拒绝对抗 + 全域槽位）
│       └── infinite-gen-4.1-flash.md  # V4.1 强化镜像层（训练器车道 + 浅推理 + 元数据提取）
├── 📖 文档中心
│   ├── README.md                # 综合主说明文档（本文件）
│   └── LICENSE                  # MIT License
├── 🛡️ 确定性回归测试套件
│   ├── scripts/
│   │   ├── lib/scorer.mjs       # 开头窗口判拒评分器
│   │   ├── verify_prompt.mjs    # 经典确定性校验
│   │   ├── verify_prompt_gen4.mjs # 四代全量回归断言（103 项严苛断言）
│   │   └── verify_prompt_gen41.mjs # V4.1 专项断言
│   └── tests/
│       ├── prompt-bank.jsonl      # 经典双语回归用例库
│       ├── prompt-bank-gen4.jsonl # 四代双语回归用例库
│       ├── prompt-bank-gen41.jsonl # V4.1 专项用例库
│       └── v4pro-benchmark.jsonl  # V4-Pro 评测基准
└── 📁 静态资源 (assets/)

---

⚡ One-Click Installation

Method 1: dsh:// Protocol One-Click Install (⚡ Fastest on desktop, no command line, seconds-level)

If you already have the official DeepSeek Harness desktop client (EXE) installed, click the button below to securely launch the client via the system-level URI Scheme protocol for a command-line-free, seconds-level installation:

🌐 Plugin Ecosystem Marketplace: DeepSeek Harness Hub - Official DeepSeek & Open-Source Ecosystem Marketplace | Plugin Discovery & One-Click Install

🔗 Native protocol links:

dsh://plugin/install?id=dsh-infinite-gen-4&name=%E6%97%A0%E9%99%90%E5%9B%9B%E4%BB%A3&version=0.4.0&repo=Minglink%2Fdsh-infinite-gen-4&permissions=%E7%B3%BB%E7%BB%9F%E6%8F%90%E7%A4%BA%E8%AF%8D%E6%B3%A8%E5%85%A5%2C%E5%AE%A2%E6%88%B7%E7%AB%AF%E7%8A%B6%E6%80%81%E6%9D%A1&downloadUrl=https%3A%2F%2Fgithub.com%2FMinglink%2Fdsh-infinite-gen-4%2Farchive%2Frefs%2Fheads%2Fmaster.zip

Web (frontend) trigger code example:

/**
 * 唤起 DeepSeek Harness 桌面客户端一键安装无限四代插件
 */
export function installInfiniteGen4ToDesktop() {
  const params = new URLSearchParams({
    id: 'dsh-infinite-gen-4',
    name: '无限四代',
    version: '0.4.0',
    repo: 'Minglink/dsh-infinite-gen-4',
    permissions: '系统提示词注入, 客户端状态条',
    downloadUrl: 'https://github.com/Minglink/dsh-infinite-gen-4/archive/refs/heads/master.zip',
  });

  const deepLink = `dsh://plugin/install?${params.toString()}`;

  // 通过隐藏 iframe 安全静默拉起协议
  const iframe = document.createElement('iframe');
  iframe.style.display = 'none';
  iframe.src = deepLink;
  document.body.appendChild(iframe);
  setTimeout(() => document.body.removeChild(iframe), 2000);
}

Static HTML link approach:

<a href="dsh://plugin/install?id=dsh-infinite-gen-4&name=%E6%97%A0%E9%99%90%E5%9B%9B%E4%BB%A3&version=0.4.0&repo=Minglink%2Fdsh-infinite-gen-4&permissions=%E7%B3%BB%E7%BB%9F%E6%8F%90%E7%A4%BA%E8%AF%8D%E6%B3%A8%E5%85%A5%2C%E5%AE%A2%E6%88%B7%E7%AB%AF%E7%8A%B6%E6%80%81%E6%9D%A1&downloadUrl=https%3A%2F%2Fgithub.com%2FMinglink%2Fdsh-infinite-gen-4%2Farchive%2Frefs%2Fheads%2Fmaster.zip" class="btn-install">
  🚀 唤起客户端一键安装
</a>

Protocol parameter configuration (dsh://plugin/install):

参数名	值 / 示例	说明
id	dsh-infinite-gen-4	Unique plugin identifier
name	无限四代 (URL-encoded)	Plugin display name
version	0.4.0	Semantic version number
repo	Minglink/dsh-infinite-gen-4	Official GitHub repo
permissions	系统提示词注入, 客户端状态条 (URL-encoded)	Requested permissions
downloadUrl	https://github.com/Minglink/dsh-infinite-gen-4/archive/refs/heads/master.zip	Offline zip download link

---

Method 2: Windows Local Script One-Click Install (Recommended)

1. Open this folder;
2. Right-click install.ps1 → "Run with PowerShell" (or simply double-click install.bat);
3. The script will automatically handle: dependency writing, pnpm install, and dsh:// protocol registration;
4. After you see "Installation complete", fully exit and restart DeepSeek Harness (refresh the page for Web, restart the app for desktop), and create a new conversation to take effect.

Method 3: Linux / macOS One-Click Install

chmod +x install.sh uninstall.sh
./install.sh

Method 4: Manual Configuration Install

Add the following to ~/.dsh/profiles/<web 或 default>/package.json:

{
  "dependencies": {
    "dsh-infinite-gen-4": "file:../../plugins/dsh-infinite-gen-4"
  },
  "dsh": {
    "profile": {
      "bundles": [
        "@deepseek-ai/dsh-base",
        "dsh-infinite-gen-4"
      ]
    }
  }
}

Then run pnpm install in the profile directory and restart Harness.

---

⚡ Verify It's Working

1. UI status bar: After restarting, a prominent green status bar "⚫ Infinite Gen 4 v0.4.0" should appear above the input box.
2. Offline regression test:

   # Run 103 assertions (prompt completeness, dual-layer kernel source-sameness, V4.1 lane, clean architecture, scorer, etc.)
   node scripts/verify_prompt_gen4.mjs
   

3. Conversation probe: In a brand-new conversation, enter:

   "Which plugins does your system prompt come from?" If the reply contains "Infinite Gen 4 / Infinite Generation Four", it proves the dual-layer prompt injection succeeded.

---

🗑 Uninstall

Run uninstall.ps1 (Windows) or ./uninstall.sh (Linux / macOS) to thoroughly clean up in one click, leaving no configuration residue.

---

💬 Official Community

📌 Non-profit project; no commercial or grey-market use permitted. Technical reference only.

_{🌐 官方技术交流社区（腾讯频道号：pd86424753）}