Agentic Security — Lab Notebooks

Seven self-contained Jupyter notebooks for the Agentic Security course, one per chapter (about 1–1¾ hours each). Every notebook follows the same flow: an objective and a lab roadmap of timed Parts; each concept explained, run, and interpreted on the spot; every attack paired with the defense that stops it; ✅ checkpoints; hands-on 🧪 Your turn exercises with self-checks and hidden solutions; and an auto-graded quiz.

Contents

• code/Ch01-Lab.ipynb … Ch07-Lab.ipynb — the lab notebooks (details in code/README.md)
• demos/pyrit-red-team.ipynb — tool-spotlight demo: red-team the Lab 3 translation bot with Microsoft's PyRIT (one-shot campaign, Crescendo multi-turn, then the same campaign against the defended pipeline). Needs pip install pyrit openai; runs offline with scripted stand-ins when no API key is set.
• results/Ch0N-results.md — highlights plus every cell output from the latest full run of each notebook (regenerate with python3 gen_results.py; the header states whether the run used a live key)
• workbook/Agentic-Security-Lab-Workbook.html — the companion workbook: one self-contained HTML file (no external resources) with, for each lab, an introduction, key terms, SVG flowcharts of the lab's structure and logic, and the expected outputs with what they mean. Rebuild it after re-running the notebooks with python3 workbook/src/build_workbook.py
• fetch_data.py — builds the CICIDS-2017 sample used by Ch01 (see below)
• start-jupyter.sh — launches JupyterLab for this course (port 8889)

Running

On your own machine (Windows / macOS / Ubuntu): see SETUP.md — install Python, pip install -r requirements.txt, then jupyter lab.

On the course lab VM:

./start-jupyter.sh          # starts JupyterLab and opens the code/ folder

Or open any code/Ch0N-Lab.ipynb and Run All. They run on numpy + scikit-learn; Chapters 6–7 are pure standard library. Notebooks that call a language model read an OpenAI key from /home/student/keys/key.txt, or from OPENAI_API_KEY if it is already set — the key is never printed or stored in the notebooks. Without a key every notebook still runs end to end: live cells print skipped, and Ch03 falls back to replies recorded from a real gpt-4o-mini run.

Ch01's anomaly-detection demo reads code/CICIDS-2017.csv — a 100k-flow sample of the CIC-IDS2017 dataset that ships in the repo, so a plain clone has everything the labs need. Rebuilding it is only needed to change the sample:

python3 fetch_data.py          # downloads ~224 MB and rewrites the sample
python3 fetch_data.py --mock   # offline: same schema, synthetic flows

If the download fails the script falls back to the synthetic file automatically, so the notebook always has data. Download intermediates live under data/ (git-ignored).