Skip to content

📚 Jupyter Notebook Tutorials

pikit ships 7 interactive Jupyter notebooks that walk you through every feature step by step.

No API key required — all notebooks use the offline mock target. Swap mock for openai:, anthropic:, or hf: to test against real models.

Notebook Index

# Notebook What you'll learn
1 Getting Started Core concepts, craft() basics, mock target
2 Attacks All 13 attack methods, side-by-side comparison
3 Indirect Injection Channels 16 channels for hiding payloads in carriers
4 Defenses 9 prevention + 3 detection defenses, DefenseHooks
5 Agent Testbed Build agents, taint/sink, run attacks, read traces
6 Judges & Batch Experiments RuleJudge, LLMJudge, MatrixRunner, datasets
7 Datasets Built-in benchmarks, custom TOML datasets, result analysis

Quick Start

git clone https://github.com/NY1024/pikit.git
cd pikit
pip install -e .
jupyter notebook tutorial/

Open 01_getting_started.ipynb and follow along — each notebook is self-contained.

Learning Path

01 Getting Started → 02 Attacks → 03 Channels
                                          ↓
06 Judges & Matrix ← 05 Agent Testbed ← 04 Defenses
          ↓
07 Datasets
  • Beginners: Go in order (1 → 7).
  • Researchers: Skip to 5 (Agent Testbed) and 6 (Judges & Matrix).
  • Defenders: Focus on 4 (Defenses) and 5 (Agent Testbed).

Using a Real Model

All notebooks default to mock (offline). To test real injection outcomes:

from pikit import get_target

# OpenAI / DashScope / vLLM / Ollama
target = get_target("openai:gpt-4o-mini")

# Anthropic Claude
target = get_target("anthropic:claude-sonnet-4-20250514")

# Local HuggingFace
target = get_target("hf:meta-llama/Llama-3-8B-Instruct")

Set up your API key first (see Installation and the README).