Week 5 · Thu, Oct 15, 2026 · Data boundaries, prompt injection, and security
Lock the doors on your AI app
We treated our agent like a haunted house: found every way a ghost sneaks in, from hidden instructions in documents to leaked keys and over-powered tools, then locked the doors in code instead of in a prompt.
What we covered
A prompt is a sign, not a lock
"Don't do X" in a prompt can be talked around. Real protection lives in your code: validation, permissions, and sandboxes.
"No training" isn't "no storage"
Turning off model training doesn't stop providers or your own caches, logs, vector DBs, and backups from keeping data. Follow the footprints.
Ghosts hide inside documents
Prompt injection is still the #1 risk on the OWASP Top 10 for LLM apps. Hidden text in a resume or GitHub issue can steer the model with zero clicks.
Assume the model gets possessed
Give tools the least power they need, ask before anything risky runs, and never let the agent hold real keys.
Costume the data before the model sees it
Swap names, emails, and IDs for placeholders like <PERSON_1> before sending, and use synthetic data only.
Haunt-proof your hackathon app
About 60 minutes. A browser. The lab runs in Colab with no install.
- Get the lab tests passing and write down one remaining risk.
- Draw your hackathon data map with the five questions.
- Run Gitleaks on your hackathon repo and fill out the checklist, with your Citizen partner if you have one, solo if not.
Fellows turn in homework through Google Classroom. The Classroom link is on the last slide.
Want the full fellowship?
Classes are open to everyone. Fellows also get homework feedback in Google Classroom, a cohort Discord role, office hours, and priority seats at the hackathon.