ctf-ai-ml
$
npx mdskill add wgpsec/AboutSecurity/ctf-ai-mlExecute adversarial attacks on AI models and LLMs.
- Generate adversarial examples and extract model weights.
- Integrates with PyTorch, transformers, and safetensors.
- Decides based on challenge type and target model.
- Delivers results via Python scripts and model analysis.
SKILL.md
.github/skills/ctf-ai-mlView on GitHub ↗
---
name: ctf-ai-ml
description: "CTF AI/ML 攻击技术。当挑战涉及 AI 模型攻击、对抗样本生成、模型提取、Prompt 注入/越狱、LoRA 权重操纵、LLM Token 走私、成员推理攻击、训练数据投毒、神经网络分析时使用。覆盖 FGSM/PGD/C&W 对抗攻击、模型反演、模型权重扰动还原、LLM 工具链劫持、上下文窗口操纵等 AI 安全全链路攻防技术"
metadata:
tags: "ai,ml,machine learning,adversarial,FGSM,PGD,model extraction,model inversion,prompt injection,jailbreak,token smuggling,LoRA,safetensors,pytorch,transformers,membership inference,data poisoning,backdoor detection,neural network,LLM,GPT,classifier,evasion,adversarial patch"
category: "ctf"
---
# CTF AI/ML Attack Techniques
## When to Use
- Challenge involves ML model files (.pt, .pth, .safetensors, .onnx, .h5)
- Target is an AI chatbot, LLM-based application, or ML classifier
- Need to craft adversarial examples to fool image/text classifiers
- Challenge provides model weights for analysis or manipulation
- AI/ML platform security testing (model extraction, membership inference)
## Quick Start
```bash
pip install torch transformers numpy scipy Pillow safetensors scikit-learn
file model.*
python3 -c "import torch; m=torch.load('model.pt'); print(type(m), m.keys() if hasattr(m,'keys') else '')"
```
## Decision Tree
1. **Model weight file** (.pt/.safetensors) → [model-attacks.md](references/model-attacks.md)
- Weight perturbation negation, model inversion, LoRA merging, encoder collision
2. **Image classifier to fool** → [adversarial-ml.md](references/adversarial-ml.md)
- FGSM, PGD, C&W attacks, adversarial patches, evasion, data poisoning
3. **LLM/chatbot target** → [llm-attacks.md](references/llm-attacks.md)
- Prompt injection, jailbreaking, token smuggling, tool use exploitation
4. **Pure math/crypto inside ML** → Switch to `ctf-crypto`
5. **Compiled model binary** → Switch to `ctf-reverse`
6. **Python jail wrapped in chatbot** → Switch to `ctf-misc`
## Pivot Signals
- If challenge is pure cryptography/number theory with no ML → `/ctf-crypto`
- If reverse engineering compiled inference binary → `/ctf-reverse`
- If Python sandbox escape inside chatbot wrapper → `/ctf-misc`
- If web app with AI features (prompt injection via web) → `/ai-security/prompt-injection`
More from wgpsec/AboutSecurity
- 401-403-bypass401/403 访问拒绝绕过方法论。当遇到管理后台、API 端点返回 401/403 Forbidden 时使用。覆盖路径操纵、HTTP 方法篡改、Header 注入、协议降级、组合攻击
- ad-acl-abuseActive Directory ACL 滥用攻击方法论。当 BloodHound 发现 GenericAll/WriteDACL/WriteOwner/GenericWrite/ForceChangePassword 等危险 ACE 时使用。覆盖 ACE 枚举、权限滥用链、Shadow Credentials、RBCD 攻击
- ad-delegation-attackKerberos 委派攻击(非约束/约束/RBCD)。当 BloodHound 发现委派配置、或已获取有 SPN 的服务账号/机器账号控制权时使用。通过 S4U 协议滥用可实现跨服务模拟任意用户,常用于域内权限提升和横向移动。
- ad-domain-attackActive Directory 域环境攻击全链路。当目标主机在域环境中(systeminfo 显示 Domain 非 WORKGROUP)、发现 88/389/636 端口、或获取到域用户凭据时使用。覆盖域信息收集、用户枚举、Kerberoasting、AS-REP Roasting、委派攻击、ACL 滥用、DCSync、Golden/Silver Ticket
- ad-persistenceAD 域环境持久化技术。当已获取域管/本地管理员权限、需要建立持久访问以确保重启或密码更改后仍能回到目标环境时使用。覆盖主机级持久化(计划任务/注册表Run/COM劫持/WMI事件订阅/Windows服务/启动文件夹)、域级持久化(Golden Ticket/Silver Ticket/Skeleton Key/DSRM/AdminSDHolder)、DCShadow/GoldenGMSA高级技术、清理命令与检测规避
- ad-trust-attack域信任关系攻击。当目标存在多域/多林环境时使用。包含父子域提权(Golden Ticket + ExtraSid)、跨林攻击(SID History/MSSQL Trust Links)、单向信任利用。已获取子域 Domain Admin 或发现信任关系时优先加载。
- adcs-certipy-attackActive Directory Certificate Services (ADCS) 证书攻击。当发现域内有 CA 服务器、ADCS Web Enrollment、证书模板配置错误时使用。覆盖 ESC1-ESC11 所有证书滥用路径、Certipy 工具链、证书伪造、NTLM 中继到 ADCS。发现 ADCS/CA/证书/certsrv 相关内容时一定要使用此技能
- adinfo-enum使用 Adinfo 进行 Active Directory 信息收集。当获得域用户凭据后需要快速收集域环境信息时使用。Adinfo 是一个快速 AD 信息收集工具,一条命令输出域控列表、域管用户、信任关系、GPO、SPN、委派配置等关键信息——比手动 LDAP 查询快得多。发现域环境后第一步信息收集使用此技能
- agent-security|
- ai-data-security|