#ai-security

Question 1: Single choice

mediumAI-checkedResponsible AI & Fairness

An attacker flips the labels of a small fraction of a spam filter's training emails, so that some legitimate emails are labelled spam. What is this attack called, and which defence addresses it most directly?

Prompt injection; filter and sanitize every single user prompt at inference time1
Data poisoning; track where training data comes from and who may change it2
Model stealing; verify and authenticate the trained model's parameters3
Overreliance; keep reminding users that the model's answers can be wrong4

Timed test