ctf-ai-ml

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.

2,140stars269forksUpdated 5/24/2026

Data Science & ML #llm-security #adversarial-ml #machine-learning #ctf #python

Security Assessment

Critical Risk(0/100)

Detected risks:

Data Exfiltration([model-attacks.md] requests.post, [llm-attacks.md] requests.post)

Remote Code Execution([model-attacks.md] eval(, [adversarial-ml.md] eval()

Secret Exposure([model-attacks.md] token =)

Sensitive File Access([llm-attacks.md] /etc/passwd)

Security Score0/100

About ctf-ai-ml

The ctf-ai-ml skill provides a focused reference for solving AI and machine learning challenges in Capture The Flag (CTF) environments. It is designed for scenarios involving ML model attacks, adversarial machine learning, large language model exploitation, and AI-specific puzzle solving. The skill helps practitioners analyze models, inspect weights and adapters, test prompt injection vectors, and identify attack paths in ML systems. It also includes guidance on when to pivot to other challenge domains such as cryptography, reverse engineering, or miscellaneous challenges when the problem no longer centers on machine learning.

FAQ

What types of CTF challenges is this skill intended for?

This skill is intended for AI and machine learning CTF challenges, including model analysis, adversarial example generation, model extraction, prompt injection, membership inference, data poisoning, LoRA adapter inspection, and LLM jailbreaking scenarios.

What tools and environment are required to use this skill?

The skill requires a filesystem-based agent environment with bash, Python 3, and internet access for installing dependencies. Supported tools include Bash, Read, Write, Edit, Glob, Grep, Task, WebFetch, and WebSearch.

Which dependencies are needed before using the skill?

The documented prerequisites include Python packages such as torch, transformers, numpy, scipy, Pillow, safetensors, and scikit-learn. Platform-specific setup may also require python3-dev on Linux or python@3 on macOS.

Does this skill support model inspection and weight analysis?

Yes. The skill includes quick-start commands and references for inspecting model formats, safetensors files, HuggingFace models, LoRA adapters, and comparing model weights between files.

Are there limitations on when this skill should be used?

Yes. If a challenge becomes primarily about mathematics, reverse engineering compiled ML binaries, or non-ML puzzles wrapped in AI interfaces, the documentation recommends switching to more appropriate domain-specific skills.