AnimeshShaw/agentic-ai-security-guide

GitHub: AnimeshShaw/agentic-ai-security-guide

面向企业高管和决策者的交互式 Agentic AI 安全治理指南，系统性覆盖威胁全景、OWASP LLM Top 10、合规框架及落地行动计划。

Stars: 1 | Forks: 0

# Agentic AI 安全指南 **面向 CISO、CIO、CTO 以及每一位需要治理具备行动能力的 AI 的高管，定制的权威性 Agentic AI 安全领导力指南。**
[![部署](https://img.shields.io/github/actions/workflow/status/AnimeshShaw/agentic-ai-security-guide/deploy.yml?label=GitHub%20Pages&logo=github&style=flat-square)](https://github.com/AnimeshShaw/agentic-ai-security-guide/actions/workflows/deploy.yml) [![CI](https://img.shields.io/github/actions/workflow/status/AnimeshShaw/agentic-ai-security-guide/ci.yml?label=CI&logo=githubactions&logoColor=white&style=flat-square)](https://github.com/AnimeshShaw/agentic-ai-security-guide/actions/workflows/ci.yml) [![安全](https://img.shields.io/github/actions/workflow/status/AnimeshShaw/agentic-ai-security-guide/security.yml?label=Security%20Scan&logo=dependabot&style=flat-square)](https://github.com/AnimeshShaw/agentic-ai-security-guide/actions/workflows/security.yml) [![0 CVEs](https://img.shields.io/badge/npm%20audit-0%20vulnerabilities-brightgreen?style=flat-square&logo=npm)](https://github.com/AnimeshShaw/agentic-ai-security-guide/actions) [![使用 Astro 构建](https://img.shields.io/badge/Astro-6.x-FF5D01?style=flat-square&logo=astro&logoColor=white)](https://astro.build) [![40 页](https://img.shields.io/badge/40%20pages-of%20content-6366F1?style=flat-square)](https://animeshshaw.github.io/agentic-ai-security-guide) [![许可证: MIT](https://img.shields.io/badge/License-MIT-yellow.svg?style=flat-square)](LICENSE)
[**在线网站**](https://animeshshaw.github.io/agentic-ai-security-guide) · [基础知识](#-foundations) · [威胁全景](#-threat-landscape) · [治理策略](#-governance--frameworks) · [阅读路径](#-reading-paths)
##

## 核心问题传统 AI 仅在被请求时才做出响应。**Agentic AI 会自主行动。** 它能够进行规划、做出决策、调用 API、执行代码、发送邮件，并在无需人工干预的情况下自主将一系列行动串联起来。当一个具备 _行动_ 能力的实体，同时面临被 prompt injection（提示词注入）、供应链攻击破坏或工具权限配置不当所操纵的风险时，其带来的安全挑战与你现有安全防御机制所能处理的范畴已截然不同。 ## 本指南旨在弥补这一空白。共 40 页深度研究内容，专为领导者编写——而非开发者。 ## 独特之处 ##

### 面向决策者，而非开发者无需死记硬背代码。没有晦涩的技术术语壁垒。每个概念都从基本原理讲起，帮助你与内部技术团队及外部供应商进行专业的沟通，并提出真正切中要害的问题。 ### 威胁驱动思维围绕真实的攻击模式构建：prompt injection（提示词注入）、工具滥用、供应链攻击、多智能体信任失效等。并非纯理论探讨——而是基于已记录的安全事件及 OWASP 的 LLM Top 10 榜单。

### 交互式学习体验包含攻防场景的翻卡片、引导式的步骤演练、具备持久化状态（可保存进度）且包含 50 个问题的高管自检清单，以及完整的可搜索术语表。 ### 生产级安全态势零 npm 漏洞。CodeQL 静态分析。Trufflehog 密钥扫描。CSP 响应头。每周 Dependabot 自动更新。这份关于 AI 安全的指南其本身就是以高度安全的方式构建的。

## 网站架构 ``` agentic-ai-security-guide/ │ ├── Foundations (3 pages) │ ├── What Is Agentic AI │ ├── How Agents Work │ └── Why Security Matters │ ├── Threat Landscape (3 pages) │ ├── Attack Surface │ ├── Prompt Injection │ └── Real-World Incidents │ ├── OWASP LLM Top 10 (11 pages) │ ├── Overview │ ├── LLM01: Prompt Injection │ ├── LLM02: Sensitive Info Disclosure │ ├── LLM03: Supply Chain │ ├── LLM04: Data Poisoning │ ├── LLM05: Improper Output Handling │ ├── LLM06: Excessive Agency │ ├── LLM07: System Prompt Leakage │ ├── LLM08: Vector Weaknesses │ ├── LLM09: Misinformation │ └── LLM10: Unbounded Consumption │ ├── Agentic AI Top 10 (11 pages) │ ├── Overview │ ├── AGT01: Prompt Injection │ ├── AGT02: Excessive Agency │ ├── AGT03: Unsafe Tool Execution │ ├── AGT04: Inter-Agent Attacks │ ├── AGT05: Memory Poisoning │ ├── AGT06: Privilege Escalation │ ├── AGT07: Uncontrolled Loops │ ├── AGT08: Data Exfiltration │ ├── AGT09: Insufficient Oversight │ └── AGT10: Supply Chain │ ├── Core Technologies (5 pages) │ ├── Model Context Protocol (MCP) │ ├── LiteLLM Proxy │ ├── NeMo Guardrails │ ├── Sandboxing Agents │ └── Tools Landscape │ ├── Governance (3 pages) │ ├── AI Governance Basics │ ├── Compliance Frameworks (NIST, EU AI Act, ISO 42001) │ └── Leadership Checklist (50 questions, saves progress) │ ├── Future Trends (2 pages) │ ├── Emerging Technologies │ └── 12-Month Action Plan (CISO/CIO roadmap) │ └── Reference #### └── Glossary（45+ 个术语，实时搜索） --- ## 阅读路径 Not every leader needs to read all 40 pages. Here are the recommended entry points: ##

Audience	Time	Path
Board Members	20 min	What is Agentic AI → Why Security Matters → Leadership Checklist
CISOs / VPs Security	3–4 hours	Threat Landscape → OWASP LLM Top 10 → Agentic Top 10 → Governance → Emerging Technologies
CIOs / CTOs	2 hours	Foundations → Core Technologies (MCP, LiteLLM, Sandboxing) → Action Plan
Risk Officers	90 min	Threat Landscape → Compliance Frameworks → Leadership Checklist
Vendor / Procurement	45 min	LLM Supply Chain → Agentic Supply Chain → Compliance Frameworks

## 基础 | Page | What You'll Learn | |------|------------------| | [What Is Agentic AI](https://animeshshaw.github.io/agentic-ai-security-guide/foundations/what-is-agentic-ai/) | How agents differ from chatbots; what autonomy really means for your enterprise | | [How Agents Work](https://animeshshaw.github.io/agentic-ai-security-guide/foundations/how-agents-work/) | LLMs, tools, memory, orchestration — the anatomy of a modern AI agent | ## | [Why Security Matters](https://animeshshaw.github.io/agentic-ai-security-guide/foundations/why-security-matters/) | 范式转变以及传统防御所遗漏的内容 | ## 威胁态势 | Page | What You'll Learn | |------|------------------| | [Attack Surface](https://animeshshaw.github.io/agentic-ai-security-guide/threats/attack-surface/) | Where agents are vulnerable and why the surface is larger than you think | | [Prompt Injection](https://animeshshaw.github.io/agentic-ai-security-guide/threats/prompt-injection/) | The most dangerous and widespread attack on AI systems, explained plainly | | [Real-World Incidents](https://animeshshaw.github.io/agentic-ai-security-guide/threats/real-world-incidents/) | Documented cases of what has already gone wrong | | [OWASP LLM Top 10](https://animeshshaw.github.io/agentic-ai-security-guide/llm-top-10/) | Full breakdown of the industry's canonical vulnerability list — all 10 entries | | [Agentic AI Top 10](https://animeshshaw.github.io/agentic-ai-security-guide/agentic-top-10/) | 10 critical risks unique to autonomous agents | | [LLM01: Prompt Injection](https://animeshshaw.github.io/agentic-ai-security-guide/llm-top-10/prompt-injection/) | The most dangerous LLM vulnerability, with interactive examples | ## | [LLM05: Improper Output Handling](https://animeshshaw.github.io/agentic-ai-security-guide/llm-top-10/improper-output-handling/) | 当 AI 输出在未经清理的情况下被信任时 | ## 治理与框架 | Page | What You'll Learn | |------|------------------| | [AI Governance Basics](https://animeshshaw.github.io/agentic-ai-security-guide/governance/ai-governance-basics/) | What you need to govern: models, agents, data, and people | | [Compliance Frameworks](https://animeshshaw.github.io/agentic-ai-security-guide/governance/compliance-frameworks/) | NIST AI RMF, EU AI Act, ISO 42001, and how they map to your obligations | ## | [Leadership Checklist](https://animeshshaw.github.io/agentic-ai-security-guide/governance/leadership-checklist/) | 每位领导者都应提出的 50 个问题 —— 附带持续进度跟踪 | ## 安全态势 This project practices what it teaches. | Control | Status | |---------|--------| | npm audit (moderate+) | Blocks CI & deployment | | CodeQL static analysis | Weekly + every PR | | Trufflehog secret scanning | Full git history on every PR | | Dependabot | Weekly updates, Monday 09:00 IST | | Content Security Policy | Strict CSP header via `_headers` | | HTTP Security Headers | HSTS, X-Frame-Options, Referrer-Policy, Permissions-Policy | ## | 依赖项中已知的 CVE | **0**（截至上次审计） | ## Tech Stack

| Technology | Role | |-----------|------| | [Astro 6.x](https://astro.build) | Static site generator — zero JS by default, island architecture | | TypeScript | Type-safe components, data schemas, and navigation config | | Custom CSS | Full design system with CSS variables, dark/light theme, zone tokens | | GitHub Actions | CI (audit → build → typecheck), Pages deployment, Security scanning | | CodeQL | Static analysis for JavaScript/TypeScript security vulnerabilities | | Trufflehog | Verified secret scanning across full git history | | Dependabot | Automated dependency update PRs | ## 本网站的制作未使用任何框架。没有 React，没有 Vue，没有 Tailwind —— 只有 Astro、TypeScript 和您能够真正看懂的 CSS。 ## CI/CD ``` 推送到 main 分支 │ ├── CI 工作流 │ ├── npm audit（中度及以上漏洞构建失败） │ ├── astro build（遇到错误即失败） │ └── tsc --noEmit（类型检查） │ ├── 部署工作流（CI 完成后执行） │ ├── npm audit（二次检查关卡） │ ├── astro build → dist/ │ └── GitHub Pages 部署 │ └── 安全工作流（同时每周运行） ├── npm audit --audit-level=high ├── CodeQL 分析 #### └── Trufflehog 密钥扫描 ## 本地开发 ``` # 克隆 git clone https://github.com/AnimeshShaw/agentic-ai-security-guide.git cd agentic-ai-security-guide # 安装 npm ci # 开发 npm run dev # 构建 npm run build # 安全审计 #### npm run audit ## 需要 Node.js 22+。 ## 贡献 Found an error? Want to add a page, expand a section, or improve the interactive elements? 1. Fork the repository 2. Create a feature branch: `git checkout -b content/your-topic` 3. Make your changes (run `npm run build` to verify) 4. Open a pull request — CI will audit, build, and type-check automatically ## 特别欢迎内容 PR、事实性更正以及新的治理案例研究。 ## 许可证 ## MIT —— 可免费使用、改编和分享。建议注明出处，但非强制要求。

Built by [Animesh Shaw](https://github.com/AnimeshShaw) · Hosted on [GitHub Pages](https://animeshshaw.github.io/agentic-ai-security-guide) *If this guide helped you think more clearly about AI security, consider starring the repo.* [![Star on GitHub](https://img.shields.io/github/stars/AnimeshShaw/agentic-ai-security-guide?style=social)](https://github.com/AnimeshShaw/agentic-ai-security-guide/stargazers)

```

标签：AI治理, Astro, CISO, OWASP LLM, 人工智能安全, 合规性, 合规指南, 静态网站