A city crossing-guard scheduler is given two risk cards: (A) the model may leak training snippets after its window is abused; (B) a planted image may push invented acceptance criteria. How should the cards be mapped?
Select an answer to reveal the explanation.
Short Explanation
Card A is about yanking secrets from training data — that is context manipulation. Card B is about tainted input that wrecks the reply into invented ACs — that is request manipulation. Keep extraction risk and disrupted-output risk on the right shelves.
Full Explanation
Official distinction: context manipulation centers on confidential training-data extraction risk, while request manipulation centers on input that disrupts output into hallucinations or other wrong testware. Swapping the labels is a common mix-up. Correct mapping drives the right tester response for each symptom.