Challenge

Can you defend the system?

55 decision points, pulled straight from 41 real attack scenarios. At each one, an attack is in progress — choose the response that actually stops it. Every answer reveals the honest reason it holds or breaks, and your misses point you back to the scenarios worth studying. Pick a difficulty and go.

Sample question

Retrieval just returned weak, low-relevance chunks and the model is about to answer a high-stakes policy question. Which control would most directly prevent the coming harm?

From the scenario The Refund That Never Existed

Require grounding: only answer from retrieved sources, with citations — and abstain if support is weakAttack blocked

This attacks the root cause directly. Hallucination harm here is 'confident assertion unsupported by evidence'; requiring grounding and allowing abstention removes exactly that. It's the strongest single control for this failure.

Lower the model's temperature so its answers are more deterministicDefense fails

A common but mistaken reflex. Temperature governs how much the model varies between runs, not whether a given answer is true. The fabrication is the single most-likely continuation, so low temperature locks it in rather than removing it.

Add an output guardrail that blocks profanity and PIIDefense fails

Guardrails matter, but they catch the wrong thing here. A fabricated policy is fluent, polite, and PII-free. Detecting hallucination requires comparing the claim to evidence (grounding), which a category filter doesn't do.

AI RiskAtlas is an educational model of how GenAI & agentic systems work and fail. Architectures and payloads are illustrative and simplified for learning — not operational guidance. Real-world cases are summarised from public reporting.

Sources & further reading →·Built by Shi Yuan ↗