Which term best describes hallucination attacks in AI systems?

• A. Attacks where the AI generates misleading, harmful, or false content.
• B. Data leakage due to weak access controls.
• C. Attacker modifies data to circumvent the model's correct prediction, causing misclassification.
• D. Attacker extracts sensitive information by analyzing the model's inputs and outputs.

Answer: (A)

Hallucinations in AI occur when the model produces content that seems plausible but is false or misleading. In this security context, a hallucination attack aims to get the AI to generate inaccurate or harmful information rather than reporting facts it knows, undermining trust and potentially causing real-world harm. This focus on the integrity and factuality of the model’s output distinguishes it from data leakage, which is about exposing sensitive data; adversarial changes that push the model to misclassify; or privacy attacks that extract information from the model’s inputs and outputs. So the best description is what happens when the AI generates misleading, harmful, or false content.