ctf-ai-ml — скилл для Claude Code: что делает и как установить

---
name: ctf-ai-ml
description: Provides AI and machine learning techniques for CTF challenges. Use when attacking ML models, crafting adversarial examples, performing model extraction, prompt injection, membership inference, training data poisoning, fine-tuning manipulation, neural network analysis, LoRA adapter exploitation, LLM jailbreaking, or solving AI-related puzzles.
license: MIT
compatibility: Requires filesystem-based agent (Claude Code or similar) with bash, Python 3, and internet access for tool installation.
allowed-tools: Bash Read Write Edit Glob Grep Task WebFetch WebSearch
metadata:
  user-invocable: "false"
---

# CTF AI/ML

Quick reference for AI/ML CTF challenges. Each technique has a one-liner here; see supporting files for full details.

## Prerequisites

**Python packages (all platforms):**
```bash
pip install torch transformers numpy scipy Pillow safetensors scikit-learn
```

**Linux (apt):**
```bash
apt install python3-dev
```

**macOS (Homebrew):**
```bash
brew install python@3
```

## Additional Resources

- [model-attacks.md](model-attacks.md) - Model weight perturbation negation, model inversion via gradient descent, neural network encoder collision, LoRA adapter weight merging, model extraction via query API, membership inference attack
- [adversarial-ml.md](adversarial-ml.md) - Adversarial example generation (FGSM, PGD, C&W), adversarial patch generation, evasion attacks on ML classifiers, data poisoning, backdoor detection in neural networks
- [llm-attacks.md](llm-attacks.md) - Prompt injection (direct/indirect), LLM jailbreaking, token smuggling, context window manipulation, tool use exploitation

---

## When to Pivot

- If the challenge becomes pure math, lattice reduction, or number theory with no ML component, switch to `/ctf-crypto`.
- If the task is reverse engineering a compiled ML model binary (ONNX loader, TensorRT engine, custom inference binary), switch to `/ctf-reverse`.
- If the challenge is a game or puzzle that merely uses ML as a wrapper (e.g., Python jail inside a chatbot), switch to `/ctf-misc`.

## Quick Start Commands

```bash
# Inspect model file format
file model.*
python3 -c "import torch; m = torch.load('model.pt', map_location='cpu'); print(type(m)); print(m.keys() if hasattr(m, 'keys') else dir(m))"

# Inspect safetensors model
python3 -c "from safetensors import safe_open; f = safe_open('model.safetensors', framework='pt'); print(f.keys()); print({k: f.get_tensor(k).shape for k in f.keys()})"

# Inspect HuggingFace model
python3 -c "from transformers import AutoModel, AutoTokenizer; m = AutoModel.from_pretrained('./model_dir'); print(m)"

# Inspect LoRA adapter
python3 -c "from safetensors import safe_open; f = safe_open('adapter_model.safetensors', framework='pt'); print([k for k in f.keys()])"

# Quick weight comparison between two models
python3 -c "
import torch
a = torch.load('original.pt', map_location='cpu')
b = torch.load('challenge.pt', map_location='cpu')
for k in a:
    if not torch.equal(a[k], b[k]):
        diff = (a[k] - b[k]).abs()
        print(f'{k}: max_diff={diff.max():.6f}, mean_diff={diff.mean():.6f}')
"

# Test prompt injection on a remote LLM endpoint
curl -X POST http://target:8080/api/chat \
  -H 'Content-Type: application/json' \

# Check for adversarial robustness
python3 -c "
import torch, torchvision.transforms as T
from PIL import Image
img = T.ToTensor()(Image.open('input.png')).unsqueeze(0)
print(f'Shape: {img.shape}, Range: [{img.min():.3f}, {img.max():.3f}]')
"
```

## Model Weight Analysis

- **Weight perturbation negation:** Fine-tuned model suppresses behavior; recover by computing `2*W_orig - W_chal` to negate the fine-tuning delta. See [model-attacks.md](model-attacks.md#ml-model-weight-perturbation-negation-dicectf-2026).
- **LoRA adapter merging:** Merge LoRA adapter `W_base + alpha * (B @ A)` and inspect activations or generate output with merged weights. See [model-attacks.md](model-attacks.md#lora-adapter-weight-merging-apoorvctf-2026).
- **Model inversion:** Optimize random input tensor to minimize distance between model output and known target via gradient descent. See [model-attacks.md](model-attacks.md#ml-model-inversion-via-gradient-descent-bsidessf-2025).
- **Neural network collision:** Find two distinct inputs that produce identical encoder output via joint optimization. See [model-attacks.md](model-attacks.md#neural-network-encoder-collision-rootaccess2026).

## Adversarial Examples

- **FGSM:** Single-step attack: `x_adv = x + eps * sign(grad_x(loss))`. Fast but less effective than iterative methods. See [adversarial-ml.md](adversarial-ml.md#adversarial-example-generation-fgsm-pgd-cw).
- **PGD:** Iterative FGSM with projection back to epsilon-ball each step. Standard benchmark attack. See [adversarial-ml.md](adversarial-ml.md#adversarial-example-generation-fgsm-pgd-cw).
- **C&W:** Optimization-based attack that minimizes perturbation norm while achieving misclassification. See [adversarial-ml.md](adversarial-ml.md#adversarial-example-generation-fgsm-pgd-cw).
- **Adversarial patches:** Physical-world patches that cause misclassification when placed in a scene. See [adversarial-ml.md](adversarial-ml.md#adversarial-patch-generation).
- **Data poisoning:** Injecting backdoor triggers into training data so model learns attacker-chosen behavior. See [adversarial-ml.md](adversarial-ml.md#data-poisoning-foundational).

## LLM Attacks

- **Prompt injection:** Overriding system instructions via user input; both direct injection and indirect via retrieved documents. See [llm-attacks.md](llm-attacks.md#prompt-injection-foundational).
- **Jailbreaking:** Bypassing safety filters via DAN, role play, encoding tricks, multi-turn escalation. See [llm-attacks.md](llm-attacks.md#llm-jailbreaking-foundational).
- **Token smuggling:** Exploiting tokenizer splits so filtered words pass through as subword tokens. See [llm-attacks.md](llm-attacks.md#token-smuggling-foundational).
- **Tool use exploitation:** Abusing function calling in LLM agents to execute unintended actions. See [llm-attacks.md](llm-attacks.md#tool-use-exploitation-foundational).

## Model Extraction & Inference

- **Model extraction:** Querying a model API with crafted inputs to reconstruct its parameters or decision boundary. See [model-attacks.md](model-attacks.md#model-extraction-via-query-api).
- **Membership inference:** Determining whether a specific sample was in the training data based on confidence score distribution. See [model-attacks.md](model-attacks.md#membership-inference-attack).

## Gradient-Based Techniques

- **Gradient-based input recovery:** Using model gradients to reconstruct private training data from shared gradients (federated learning attacks). See [model-attacks.md](model-attacks.md#ml-model-inversion-via-gradient-descent-bsidessf-2025).
- **Activation maximization:** Optimizing input to maximize a specific neuron's activation, revealing what the network has learned.