Lab 05: RAG Adversarial Harness

Series: AI Eval Labs · Phase: Robustness and Security · Depends on: Lab 03

Core scope

• [ ] Inject irrelevant context and contradictory documents
• [ ] Remove citations and test for grounded abstention
• [ ] Score confident-but-ungrounded answers as failures

Stretch

• [ ] Per-perturbation degradation curves

Links

• Code: series/ai-eval/05-rag-adversarial
• Demo: coming soon
• Write-up: coming soon on markrox.ai