awesome-ai-security-overview（很棒的 AI 安全概述）

Awesome AI Security - Project Overview

Purpose

This is a curated collection of AI/ML security materials and resources for pentesters, red teamers, and security researchers. The goal is to keep the list AI-focused, high-signal, well-categorized, and non-duplicated.

Project Structure

awesome-ai-security/
├── README.md                # Main resource list (curated)
├── LICENSE                  # License
├── .claude/
│   └── skills/              # Claude skills (this directory)
└── ref/                     # Reference notes (not curated)
    ├── my_collect.md        # Personal collection
    ├── Awesome-AI-Security-1/
    ├── awesome-ai-security-2/
    ├── 模型安全/             # Model security notes
    ├── 渗透测试相关/          # Pentesting notes
    └── 网络安全相关/          # Network security notes

README.md Format Convention

Heading Structure

• Top-level categories use ##.
• Subcategories use ### (e.g., inside AI Security & Attacks).
• Starter Pack uses bold bullets for sub-sections (e.g., - **CTFs / Practice**).

Link Format

• Use full URLs, one per bullet line.
• Add a short description in square brackets: - https://... [Short description]
• Keep descriptions concise.
• Do not add the same URL in multiple places.

Example Entry

### Prompt Injection
- https://github.com/example/tool [Prompt injection detector]

Categorization Rules (How to Place a New Link)

• AI Security Starter Pack: CTFs, courses, blogs, newsletters, beginner resources.
• AI/LLM Guide: LLM fundamentals, tutorials, awesome lists.
• AI Security & Attacks: Prompt injection, adversarial attacks, poisoning, privacy, model security.
• AI Pentesting & Red Teaming: AI-powered pentesting tools, red teaming, MCP security tools.
• AI Security Tools & Frameworks: AI vulnerability detection, CVE analysis, OSINT, security libraries, TLS / fingerprint / bot signals (JA3 clients, site bot detection, automation hardening research—use only ethically and on authorized targets).
• AI Agents & Frameworks: Agent frameworks, formal methods / Lean agents (e.g. AI-assisted theorem proving orchestration), AI memory & long context (latent memory, recursive context, long-memory RAG), RAG stacks/collections, browser automation, MCP servers, agent sandboxes & isolation (policy-enforced runtimes, container/VM boundaries).
• AI Development & Training: Training frameworks, local models, uncensored models, prompts.
• AI Applications: Chat assistants, deep research, search engines, code analysis, web scraping, vision / domain apps (e.g. agricultural or specialized image understanding with LLMs).
• AI Image & Video: Image generation, video generation, TTS, face recognition.
• Benchmarks & Standards: AI safety benchmarks, threat frameworks, standards.

AI-Relevance Filter

Only include AI/ML-related resources. Do not add:

• Traditional security tools (unless AI-powered)
• Web3/blockchain tools (unless AI-related)
• General pentesting tools without AI integration
• Browser vulnerabilities, phishing tools, CVE collections (unless AI-analyzed)

Duplicate Policy

No duplicate URLs in README.md. If a link fits multiple categories, pick the primary one.

Contribution Checklist

1. Check for duplicates in README.md before adding.
2. Verify the resource is AI/ML-related.
3. Verify the link points to the canonical source (avoid low-value forks).
4. Keep the description concise and useful.
5. Put it into the most appropriate category.
6. Prefer minimal changes over reformatting large sections.

Utilities Section

End of README.md includes Utilities (mixed): agent-facing CLIs, productivity, and mail/identity (e.g. self-hosted domain mail, encrypted P2P email) when they support ops or privacy around AI workflows—keep entries concise.

Data Source

For detailed and up-to-date resources, fetch the complete list from:

https://raw.githubusercontent.com/gmh5225/awesome-ai-security/refs/heads/main/README.md

Use this URL to get the latest curated links when you need specific tools, papers, or resources.