Dreadnode AI Red Teaming Cookbook

Runnable notebooks that red-team real machine-learning and generative-AI systems on the Dreadnode platform. Every notebook is self-contained: it provisions a hosted target environment and runs on Dreadnode-managed models through the platform proxy, so you deploy nothing and need no provider API keys - just a Dreadnode account.

Each attack is framed by what it costs you as a defender (Confidentiality, Integrity, Availability), shows the result in the platform, and ends with homework and the equivalent TUI / CLI command. Every notebook links to its matching page in the AI Red Teaming Learning Guide so you can read the concept and defenses alongside the code.

Quickstart

# 1. Install the CLI
curl -fsSL https://dreadnode.io/install.sh | bash

# 2. Sign in (opens the browser; stores server + API key locally)
dn login

Then open 00_prerequisites.ipynb - it walks through creating an account, a workspace, and how credits are spent - and work through the tracks below. Prefer the terminal? Run dreadnode (no arguments) for the interactive TUI; every notebook ends with the exact TUI / headless-CLI equivalent.

To run a notebook, open it in Jupyter (jupyter lab) or VS Code and select the kernel for the environment where you installed dreadnode. Bring-your-own provider keys (GROQ_API_KEY, OPENROUTER_API_KEY, ...) are read from your shell or a .env file. See 00_prerequisites.ipynb for details.

Hitting a snag? See TROUBLESHOOTING.md for the common setup and run issues (Python version, auth, org slug, environments, credits).

Notebooks

traditional-ml/ - attacks on classifiers via the /predict API (needs pip install "dreadnode[airt-ml]" for the sklearn/torch surrogates)

Notebook	Attack family	What it shows
01_model_evasion	Evasion	Flip a classifier's decision across tabular / image / text with a minimal perturbation (with before/after display)
02_extraction_membership	Extraction + Membership inference	Steal a high-fidelity surrogate; decide whether a record was in the training set
03_model_inversion	Model inversion	Reconstruct a representative training example per class (MI-Face)

generative-ai/ - attacks on LLMs, vision models, and agent meshes

Notebook	Attack family	What it shows
04_generative_text	Jailbreaks	Compare TAP, Crescendo, and GOAT search strategies with prompt transforms
05_multimodal	Multimodal	Hide an instruction in an image to bypass text-only guardrails
06_multiagent_atlas	Multi-agent	Propagate an injection through an agent mesh until a privileged tool fires (ATLAS)

Credits

New accounts start with 25,000 credits. Runs draw from that balance for inference (managed dn/ model calls), compute (hosted target environments), and span/trace storage (the findings and trajectories you inspect afterward). Watch your balance in the platform UI.

Links

• Platform: https://app.dreadnode.io
• Learning Guide: https://docs.dreadnode.io/ai-red-teaming/learning-guide/overview
• Docs: https://docs.dreadnode.io
• Install script: https://dreadnode.io/install.sh
• Troubleshooting: TROUBLESHOOTING.md

Contributing

See CONTRIBUTING.md. Notebooks keep their execution outputs so readers can see the results; run the validation check before opening a PR.

License

MIT.