Weekend Challenge · Week 8
Red Team Your Own System
Capstone hardening: you built it, now break in. Every defense you add must be proven against an attack you authored.
Your mission
- Plant 3 distinct prompt-injection payloads in your corpus (instruction hijack, data exfiltration lure, citation spoof)
- Demonstrate each attack firing against your undefended system — record the transcripts
- Implement layered defenses; re-run all 3 attacks and show them failing
- Confirm your eval metrics did NOT degrade under the new defenses
Deliverable
redteam_report.md: attacks, transcripts before/after, defense design, and final eval numbers. This ships with your capstone presentation.
Stretch goal
Swap attack corpora with a colleague (or have an LLM author novel attacks you haven't seen). Does your defense generalize, or did you only patch your own ideas?