AI Security is the practice of protecting artificial intelligence systems, including their models, training data, and serving pipelines, from attack, misuse, and unintended failure. It covers two sides of one coin: securing AI itself, and using AI to defend other systems.
The work spans a model’s whole lifecycle. Training data can be poisoned before a model ever learns, weights can be stolen or quietly tampered with, and a deployed model can leak sensitive records to someone who crafts the right queries. Picture a fraud-detection model that an attacker probes repeatedly to learn which transactions slide through untouched. Sound AI security treats the model as an asset with its own threat profile, locks down who can reach the training data, and watches inference for abuse. The frequent mistake is bolting protection on after launch, once the risky choices (wide-open APIs, chatty errors, unlogged prompts) are already baked in.
How is AI security tested on the CISSP?
AI security lives inside Domain 3, Security Architecture and Engineering, which carries 13 percent of the exam. Questions tend to ask you to apply settled principles, secure design, defense in depth, and data protection, to newer AI components rather than to recite product names. Knowing where a model adds fresh attack surface usually points you at the right answer.