AI/ML CTF Quick Reference

Problem

AI/ML CTF challenges often treat model weights, inference APIs, LLM agents, and training data as attack surface. Solvers may struggle to distinguish model weight analysis, adversarial ML, and LLM prompt or tool abuse, especially when the challenge looks like a normal chatbot, a fine-tuned model, or a protected classifier. This skill gives a concise reference for choosing the right technique and avoiding false pivots.

Core capabilities and workflow

• Model weight analysis: for fine-tuned models, LoRA adapters, or inference endpoints, try negating fine-tuning deltas with 2*W_orig - W_chal, merge adapters as W_base + alpha * (B @ A), or use gradient descent for model inversion and encoder collision.
• Adversarial ML: for classifiers or vision models, evaluate attacks in escalation order: FGSM, PGD, C&W, adversarial patches, data poisoning, and backdoor detection; physical-scene tasks often favor patches and triggers.
• LLM attacks: for chatbots, RAG systems, and agents, inspect direct and indirect prompt injection, jailbreaking, token smuggling, context window manipulation, and tool use exploitation.
• Boundary checks: pivot to crypto for pure math, lattice reduction, or number theory; to reverse for ONNX, TensorRT, or compiled inference binaries; to misc for ML-wrapped games or puzzles.