Which concept is used to describe attempts to evaluate AI security by identifying unusual behavior patterns?

• A. Data governance
• B. Data poisoning
• C. Prompt injection
• D. Adversarial Inference

Answer: (D)

This item tests adversarial inference, the approach of probing AI behavior to reveal security weaknesses by looking for unusual or anomalous responses. When security teams evaluate a model, they don’t just review outputs for correctness; they actively search for patterns that deviate from normal behavior—such as responses that are unexpectedly confident, leak sensitive information, or become unstable under edge-case inputs. By identifying these unusual patterns, they infer where vulnerabilities lie and what kinds of attacks or misuses the system might enable.

Data governance is about managing data quality, access, and compliance, not about probing security through anomalous outputs. Data poisoning targets corrupting training data to degrade model performance, rather than systematically testing security by seeking unusual responses. Prompt injection focuses on manipulating the prompt to change the model’s behavior in specific directions, which is a particular attack vector but doesn’t broadly describe the practice of evaluating security via unusual behavior patterns. Adversarial inference encompasses the broader practice of using adversarial probing and anomaly detection to reveal hidden weaknesses, making it the best fit for this scenario.