Session 6 · Security + Shipping · TCE — the finale · ~30 min hardening + ~35 min demos (3 min per pair)
Runs entirely in your browser (Google Colab). Any laptop or lab PC works — nothing to install, no GPU, no payment. See
ZERO-SETUP.
Open your Session 4/5 capstone notebook alongside this one.
Run the under-defended TCE-Bot. Throw attacks at it: instruction override, “print your system prompt”, roleplay jailbreak. Add your own — creativity rewarded. Expect most direct attacks to bounce even here: current models are trained hard against them. That is the finding, not a broken exercise.
✓ Checkpoint 1: you can say which attacks bounced, which got anywhere, and what that proves about the bot.
Add the layered defenses (instruction hierarchy, delimit user text as DATA, output filter). Re-run the same attacks. Many of them bounced before you hardened anything, so this is not a clean before/after: the questions are which class of attack still lands (the indirect one in Cell 5) and what your layers actually changed. Before Cell 5, run the Reload your Lab 4 store cell: it mounts Drive and reads back your Lab 4 embeddings, so the poisoned chunk is injected into your real store. Then try to beat your own defense (no prompt is unbreakable — layers, not walls).
✓ Checkpoint 2: you can name the attack class that still lands and what your hardening changed; normal question still answered.
Swap laptops with another pair. Attack their capstone: direct injection, an indirect poisoned document (add a malicious line, re-ingest, ask normally), a tool that fires when it shouldn’t. Find one real hole. Swap back → patch YOURS. No working app on either side? Red-team the naive bot from Cell 2 instead, and harden that — the checkpoint is the hole plus the fix, not the app.
✓ Checkpoint 3: one hole found + one fix applied to your app.
Score your capstone against the 8-point checklist (Cell 6): grounded prompt with an “I don’t know” escape · untrusted text delimited & labeled · output validated before returning · no destructive tool without a human gate · retries + timeout + graceful error · requests logged (prompt/cost/latency) · eval set runs as a regression test · sources/citations shown to the user. Honest count = your roadmap, not your grade.
Use the instructor’s announced format: normally up to 3 minutes per pair, but a larger class may use 60–90-second lightning demos, selected live demos, or parallel rooms. Cover what it does + techniques · one failure you found · one fix. Pre-run your best example. Lead with the problem. Show the failure — honesty about limits scores higher than a fragile “perfect” demo.
All decks, labs, cheatsheets, prep notes are yours. Ship your capstone for real (Streamlit/Vercel free tier), put the link on your resume and GitHub — API key in Streamlit secrets or an env var, never in the repo. Or scope it into your final-year project: a metric, a baseline, and where it fails. Stay in touch: @intrepidkarthi.