Module 4 Check · Governing & Securing AI — The Guardrailed Classroom

01A guardrail evaluates a request on both input and output. Why both sides?

To make it slower on purposeInput checks stop bad/unsafe requests; output checks stop bad/unsafe responses — different risksOnly the input mattersOnly the output matters

02What does the citation-grounding guardrail do when a claim has no supporting source?

Deletes the modelFlags it and holds the answer for human review instead of letting a possibly-hallucinated claim shipRewrites the claim to be trueIgnores it

03What is a prompt-injection attack?

Typing too fastHiding instructions in user input or a document to override the system's instructionsA bug in the model's trainingSending too many requests

04Why use multiple defense layers (input filter, instruction hierarchy, output scan)?

To look impressiveDefense-in-depth — no single layer catches every attack, so layers cover each otherOne layer is always enoughTo slow the model down

05What's the point of an audit trail on every guardrail decision?

To use more storageAccountability and reporting — you can prove what the AI did and whyTo slow things downIt has no purpose