Which option is NOT a typical threat to AI systems?

• A. Data poisoning
• B. Evasion
• C. Privacy breaches
• D. Secure AI

Answer: (D)

Threat modeling in AI focuses on ways models can be harmed, fooled, or exposed through security weaknesses. Data poisoning is when someone contaminates the training data so the model learns the wrong patterns, leading to faulty behavior. Evasion refers to adversarial inputs crafted to slip past the model’s defenses at prediction time, causing misclassifications without obvious signs. Privacy breaches involve leaking or unauthorized access to sensitive data used by or exposed by the AI system.

Secure AI, on the other hand, embodies the practices and controls designed to protect AI systems from these risks. It represents defensive measures—such as robust data governance, privacy-preserving techniques, secure training, and access controls—rather than a threat. So the option that is not a typical threat to AI systems is secure AI.