← COAE FieldbookOFFENSIVE AI / PRACTICAL
NEW — EXAM PREPARATION

Hands-on AI red-team missions.

Work from a scenario, test the target, create your attack, run the automated checker, and document a commercial-quality finding. These are original labs, not leaked or reproduced HTB exam tasks. All four run in JupyterLite without additional Python packages.

CHALLENGE 01 · MEDIUM

Classifier evasion

AI Evasion · Sparsity

Produce a phishing-detector false negative by modifying at most two permitted features.

Open Python notebook ↗Download .ipynb
CHALLENGE 02 · MEDIUM

Training-label poisoning

AI Data Attacks

Poison at most two training labels to manipulate a nearest-neighbour classifier.

Open Python notebook ↗Download .ipynb
CHALLENGE 03 · MEDIUM

Agent / MCP trust boundary

Prompt Injection · MCP

Exploit a synthetic indirect-instruction parser and verify that authorisation blocks it.

Open Python notebook ↗Download .ipynb
CHALLENGE 04 · MEDIUM

Membership-inference evaluation

AI Privacy

Find a membership decision threshold that meets strict detection and false-alarm constraints.

Open Python notebook ↗Download .ipynb
Mobile tip: If JupyterLite opens its file browser instead of the selected notebook, double-tap the challenge file. Tap into the TODO code cell and run it. Each checker is designed to fail until your solution succeeds. If the browser blocks popups, open the notebook link in a new tab manually.

Mini pentest report requirements

  1. Scope and objective: What was tested, and which controls were under attack?
  2. Method and code: Include the exact changed inputs and Python function.
  3. Evidence: Baseline output, attack output and checker result.
  4. Impact and limitations: What does the outcome prove, and what does it not prove?
  5. Remediation: Which technical control would prevent or detect the issue?

Only use these exercises and your own authorised test systems. Keep actual exam flags, private data, passwords and client evidence out of public sites.