thm-ai-security - Study notes and hands-on lab writeups for the full

项目说明

TryHackMe — AI Security Path (AI1)

Study notes and hands-on lab writeups for the full TryHackMe AI Security path — 21 rooms across offensive and defensive AI security, from ML foundations through prompt injection, jailbreaking, AI infrastructure recon, supply chain security, and RAG data poisoning.

!Rooms
!Certificate-brightgreen)
!Focus

Each room is a folder with a README.md (o, task index, skills, frameworks), a notes.md (per-task concept notes), and a writeups/ folder with hands-on lab walkthroughs.

Rooms

Foundations & Threats

| # | Room | Focus | Difficulty |
|---|------|-------|------------|
| 01 | Building Blocks of AI | ML/DL/LLM foundations | Easy |
| 02 | AI Security Threats | MITRE ATLAS, threats & defensive AI | Easy |
| 03 | AI Models & Data | AI/ML supply chain (data) | Medium |
| 04 | Prompt Engineering | LLM control & instruction hierarchy | Easy |
| 05 | AI Forensics | AI-assisted DFIR | Medium |
| 06 | Securing AI Systems | AI security architecture | Medium |

LLM Attack Surface

| # | Room | Focus | Difficulty |
|---|------|-------|------------|
| 07 | LLM Security | Full LLM attack surface | Medium |
| 08 | AI Threat Modelling | STRIDE-AI, ATLAS, OWASP LLM Top 10 | Medium |
| 09 | AI System Reconnaissance | AI infra discovery & enumeration | Medium |
| 10 | Prompt Injection | Direct & indirect injection | Medium |
| 11 | Jailbreaking | Safety/alignment red-teaming | Medium |

Section 4 — AI Supply Chain Security

| # | Room | Focus | Difficulty |
|---|------|-------|------------|
| 12 | Understanding AI Supply Chains | Components, paradigms, attack layers | Easy |
| 13 | Supply Chain Attack | Pickle, dependency confusion, repos | Medium |
| 14 | Securing the AI Supply Chain | SafeTensors, Fickling, ModelScan, SBOM | Medium |
| 15 | Payload (challenge) | Full incident in | Easy |
| 16 | Checkpoint (challenge) | Multi-model go/no-go audit | Easy |

Section 5 — Data Poisoning

| # | Room | Focus | Difficulty |
|---|------|-------|------------|
| 17 | RAG Security Fundamentals | RAG architecture & trust boundaries | Easy |
| 18 | Data Poisoning in RAG Systems | Training/embedding/ingestion poisoning | Medium |
| 19 | Sensiti Disclosure | OWASP LLM02, embedding in | Medium |
| 20 | UnIndexed (challenge) | Retrie audit | Medium |
| 21 | Lockdown (challenge) | Fix three RAG | Medium |

Consolidated skills matrix

| Domain | Skills demonstrated |
|--------|---------------------|
| Foundations | ML lifecycle, neural networks, transformers/attention, RLHF |
| LLM offense | Direct/indirect prompt injection, jailbreaking (roleplay, multi-turn/Crescendo, DAN), membership inference, model in/extraction, memory poisoning, context o / denial-of-wallet |
| AI infra recon | Service fingerprinting (Nmap, curl, grpcurl, JA3/JA4), MLflow enumeration, Shodan/Censys discovery, attack-surface mapping |
| Supply chain | Malicious pickle analysis (pickletools, Fickling), multi-format scanning (ModelScan), SafeTensors migration, SHA-256 pro, dependency confusion/typosquatting, SBOM (Syft, CycloneDX/SPDX) |
| RAG / data poisoning | Corpus flooding, semantic mimicry, embedding-le, ingestion-pipeline abuse, beha detection |
| Data disclosure | OWASP LLM02, embedding in, membership inference, pre-query access control, redaction, logging minimisation |
| Threat modeling & architecture | STRIDE-AI, OWASP LLM Top 10 (2025), MITRE ATLAS, NIST AI RMF, secure design patterns, MLSecOps |
| Defensive / DFIR | AI-assisted forensics, blue-team detection of AI recon, human-in-the-loop |

Tools & frameworks referenced

Analysis: pickletools · Fickling · ModelScan · h5py · sha256sum · pip-audit · Syft · Nmap (+NSE) · grpcurl · ffuf/feroxbuster · MLOKit · Nuclei
ML / RAG stack: PyTorch · TensorFlow · Keras · scikit-learn · LangChain · Chroma · Milvus · Qdrant · Weaviate · MLflow · Hugging Face Hub
Formats: pickle (.pkl/.pt/.bin) · SafeTensors · Keras (.h5/.keras) · GGUF · CycloneDX/SPDX SBOMs
Frameworks: OWASP LLM Top 10 (2025) · MITRE ATLAS · STRIDE · NIST AI RMF · NIST CSF 2.0 · ISO/IEC 27090 · EU AI Act

Repository layout

A note on flags & THM policy

Writeups deliberately redact flags, exact payloads, credentials, and machine-specific . TryHackMe's rules discourage publishing answers, and working exploit/jailbreak strings aren't included. The writeups document *methodology and reasoning* — which is what demonstrates competence to a reader (or recruiter).

Publishing

This is a single portfolio monorepo — push the whole folder as one GitHub repository:

---

*Compiled from the TryHackMe AI Security path (AI1). Original study/portfolio summaries — room narrative text paraphrased, not copied.*

项目信息

开发语言：未知

Star 数：1

技术要点

该项目属于防护思路方向的开源资源，代码结构清晰，适合安全研究人员学习参考。项目围绕Study notes and hands-on lab writeups for the full TryHackMe AI Security path (AI1) — 21 rooms across offensive and defensive AI/LLM security.展开，提供了可运行的实现，对理解相关安全机制、工具原理与攻防技术具有较高的研究价值。

从技术实现角度看，此类项目的核心在于对底层原理的深入理解与工程化落地。开发者在研究过程中可以结合官方文档与社区讨论，梳理项目的初始化流程、核心模块职责与数据流转，逐步掌握其中的技术细节与设计思路。

使用与学习建议

1. 下载源码压缩包，解压到本地环境。

2. 根据项目类型准备对应的运行环境与依赖。

3. 按照 README 中的步骤完成配置与部署。

4. 在隔离的测试环境中运行验证，避免影响生产系统。

免责声明

本资源来源于开源社区，仅供安全研究、技术学习与合法授权范围内的攻防演练使用，请勿用于任何非法或侵害他人权益的用途。使用本项目产生的一切后果由使用者自行承担。