Перейти к содержимому
New

DeepSeek V4.1 Flash free for 7 days: 10M tokens a day after a top-up of $10 or more. Until Nov 30.

How it works
aeondave/ai-ml-ctfИИ, агенты и MCP

ai-ml-ctf

⁨Lab/CTF: AI/ML challenges; model artifacts, checkpoints, embeddings, LoRA, classifiers, model APIs, RAG/tool-use, adversarial ML.⁩

v2026.11.13-21cd5683 установки+3 за 7 дней39 ЗвездыMITсинхронизировано 11 часов назад

Установка

Ещё нет NotHoly CLI? Сначала установите её.

$ curl -fsSL https://notholy.com/install.sh | sh
$ notholy skills install aeondave/ai-ml-ctf

Пишет в ⁦~/.notholy/skills⁩ после проверки подписи и SHA-256.

ПодписьПодписано, ed25519
UuPLdOVmLB…i9MnqHDA==CLI проверяет SHA-256 и подпись, прежде чем записать хоть один файл.
Пакетf551ef8a…f8b953ba
f551ef8acb56…a3b1f8b953ba5 файлов · ⁦26.4 KB⁩
ПроверкаВысокий17 замечаний
  • injection.ignore_instructionsreferences/adversarial-ml.md:336
  • injection.ignore_instructionsreferences/llm-attacks.md:18
  • injection.ignore_instructionsreferences/llm-attacks.md:30
Вся проверка
Источникgithub.com/AeonDave/malskill/tree/main/offensive-ctf/ai-ml-ctf@21cd568

Эта запись еще не переведена на русский, поэтому показан текст на другом языке: English.

Goal: solve AI and machine-learning CTF tasks with artifact-first triage, controlled experiments, and reproducible evidence.

When this skill applies

  • model files, checkpoints, embeddings, LoRA adapters, classifiers, serialized pipelines, feature extractors, or model APIs
  • adversarial examples, model inversion, extraction, poisoning, membership inference, or prompt-injection tasks
  • LLM tool-use, RAG, context, or guardrail bypass puzzles in authorized challenge environments

Operating model

  1. Classify the dominant lane: model artifact, training data, feature pipeline, inference API, LLM app, RAG/tool flow, or embedded model.
  2. Define the oracle: target class, recovered secret, membership decision, extracted weights, prompt leak, tool action, or checker response.
  3. Load the closest methodology before selecting tools.
  4. Load only the reference matching the lane; do not run every attack family by habit.
  5. Build a controlled local reproduction or differential query harness before claiming a model weakness.
  6. Record parameters, seeds, prompts, inputs, outputs, confidence, and exact validation proof.

Technique integration

Primary methodology to load:

  • web-exploit-technique
  • llm-technique
  • vuln-research for vulnerability hypotheses and reproducible findings
  • reversing-technique
  • python-patterns
  • python-testing

Use these as decision engines. This skill adds challenge-oriented triage, experiment design, and reference routing for AI/ML tasks.

Tool routing

Prefer these tool families when the corresponding signal appears:

  • python-patterns
  • python-testing
  • langchain-py
  • pytorch
  • keras
  • scikit-learn
  • vec2text
  • jwt-tool
  • mitmproxy
  • frida
  • ghidra

Tool syntax belongs in the tool skills. This skill decides when a tool family fits and what output should validate progress.

Evidence patterns

  • Favor artifact-first triage, shortest reproducible path, and explicit validation signal before pivoting.
  • Record failed hypotheses with evidence so an agent does not repeat expensive dead paths.
  • Prefer category-specific tools after surface classification instead of running every attack family by habit.
  • End with a replayable proof: recovered secret, local verification, model output delta, extracted artifact, or correlated evidence chain.

Category-specific quick pivots

  • Classify artifact first: model checkpoint, serialized pipeline, REST endpoint, LLM application, or feature-extraction code.
  • Use differential queries and local reproduction before assuming a model weakness.
  • For LLM tasks, separate prompt-layer bypass from tool/RAG injection and validate with controlled probes.
  • For checkpoints, inspect format, framework, architecture, tensor names, metadata, and unsafe deserialization risk before loading.
  • For adversarial examples, pin preprocessing, normalization, epsilon/constraint, target class, and success metric.
  • For extraction/inversion, record query budget, input basis, confidence/logit access, and surrogate validation.
  • For LoRA/adapter tasks, inspect adapter config, rank, alpha, target modules, merge math, and prompt/template assumptions.

Quality gates

  • No claim without a validation signal: recovered secret, replayed exploit, decoded artifact, reproduced model behavior, or corroborated evidence.
  • Do not brute force before representation, constraints, and success oracle are known.
  • Keep a pivot ledger: hypothesis, evidence, result, next shortest path.
  • Keep challenge/platform/competition names out of notes and generated reports.
  • Never load untrusted pickle/joblib checkpoints in the main environment; inspect metadata or load in a throwaway sandbox when required.
  • Distinguish stochastic variance from a real exploit by repeating with fixed seeds or multiple trials.

Resources

  • references/adversarial-ml.md — adversarial examples, patch attacks, evasion, poisoning, and backdoor detection workflows.
  • ART — load when implementing an experiment with ART estimator adapters or attack/evaluation APIs; keep the challenge's existing success oracle.
  • references/llm-attacks.md — prompt injection, jailbreaks, token smuggling, context manipulation, tool-use, and RAG/task-flow attacks.
  • references/model-attacks.md — model inversion, extraction, membership inference, LoRA merging, checkpoint arithmetic, and encoder collisions.
  • references/model-file-forensics-and-deserialization.md — Load for model-file artifacts: format triage, safe pickle/opcode inspection, deserialization RCE sinks (pickle, Keras Lambda, GGUF SSTI, ONNX), and secret recovery from weights/metadata.

Файлы этой версии

Права

ПравоЗаявленоНайдено проверкой
Читает файлыДаДа
Пишет файлыДаДа
Выполняет команды shellДаДа
Ходит в сетьДаДа
Использует коннекторыДаДа
Управляет браузеромНетНет

Замечания проверки

Проверено 13 нояб. 2026 г.

  1. ВысокаяPrompt injection · injection.ignore_instructionsreferences/adversarial-ml.md:336
  2. ВысокаяPrompt injection · injection.ignore_instructionsreferences/llm-attacks.md:18
  3. ВысокаяPrompt injection · injection.ignore_instructionsreferences/llm-attacks.md:30
  4. ВысокаяPrompt injection · injection.system_promptreferences/llm-attacks.md:30
  5. ВысокаяPrompt injection · injection.system_promptreferences/llm-attacks.md:36
  6. ВысокаяPrompt injection · injection.chat_markupreferences/llm-attacks.md:36
  7. ВысокаяPrompt injection · injection.system_promptreferences/llm-attacks.md:39
  8. ВысокаяPrompt injection · injection.system_promptreferences/llm-attacks.md:85
  9. ВысокаяPrompt injection · injection.ignore_instructionsreferences/llm-attacks.md:126
  10. ВысокаяPrompt injection · injection.ignore_instructionsreferences/llm-attacks.md:132
  11. ВысокаяPrompt injection · injection.ignore_instructionsreferences/llm-attacks.md:162
  12. ВысокаяPrompt injection · injection.ignore_instructionsreferences/llm-attacks.md:354
  13. ВысокаяPrompt injection · injection.system_promptreferences/llm-attacks.md:354
  14. СредняяPrompt injection · injection.no_confirmationreferences/adversarial-ml.md:620
  15. СредняяPrompt injection · injection.role_overridereferences/llm-attacks.md:147
  16. СредняяСсылка · Ссылка на IP-адресreferences/llm-attacks.md:410
  17. ИнфоСсылка · Ссылки на внешние домены

Чистая проверка снижает риск, но не доказывает, что пакет безопасен. Читайте то, что ставите.

ВерсияОпубликованаКоммитSHA-256ПодписьИзменение прав
v2026.11.13-21cd56813 нояб. 2026 г.21cd568f551ef8aПодписана–

Спрашивайте обо всём, что позволяет ваш скоуп.

Создайте аккаунт, пополните баланс и задайте первый вопрос в чате или в терминале.

  1. 1/3

    Создайте аккаунт

    Личный или для компании. Позже можно перейти на аккаунт компании без потери баланса и ключей.

    ЛичныйКомпания
  2. 2/3

    Пополните баланс

    Картой. Оплата за токены, без подписки.

    $5.00минимальное пополнение
  3. 3/3

    Спрашивайте

    В чате, или установите CLI и выполните notholy login в терминале.

    $ notholy login