
Hiding sensitive values inside an agent's reply without discarding the rest of the answer depends on which level of Sensitive Data Protection is configured in Model Armor. Model Armor offers two levels, basic and advanced, and the boundary between them determines whether sensitive values are only flagged or actually masked before a response reaches a user. For the Professional Agentic Architect exam, this distinction matters because several other Model Armor controls can detect violations or block output, yet only advanced Sensitive Data Protection masks individual matches while letting the rest of a response go through.
Basic Sensitive Data Protection supports inspection only, and it works solely with predefined infoTypes. It can flag values such as credit card numbers or US Social Security numbers in an agent's output, but it does not mask them. Even when basic inspection flags a match, the full value still reaches the user. Basic Sensitive Data Protection also cannot recognize custom formats, so an internal ID format falls outside what it can detect.
Advanced Sensitive Data Protection references Sensitive Data Protection templates instead of stopping at predefined inspection. This configuration pairs two templates that serve distinct roles:
When configuring advanced Sensitive Data Protection, every infoType listed in the de-identify template must also appear in the inspect template. Pairing both templates lets an agent keep delivering the useful portion of its answer while hiding the sensitive values inside it, whether those values are predefined formats like credit card numbers and US Social Security numbers or custom internal identifiers.
Advanced Sensitive Data Protection is a deterministic control. Detection and masking follow fixed patterns rather than model behavior. A custom regex infoType matches text against a regular expression, so a value in a given format is detected by pattern rather than by judgment, the same way each time. The de-identify template then masks each match instead of only reporting it, and the rest of the text passes through.
Controls that work through a model behave differently. Prompt instructions, sampling settings such as temperature, and few-shot examples influence what a model writes, and safety classifier models such as Llama Guard judge content with a model of their own. Their results depend on model behavior, so they can make a sensitive value less likely to appear without guaranteeing that it never does. Blocking a whole response after a classifier flags it is also different from masking one value inside it.
Sensitive Data Protection is one part of a Model Armor template, and the settings around it each have a distinct job. Knowing what each one does makes it clear that de-identification is the setting that masks part of a response.
Masking therefore comes from the advanced Sensitive Data Protection configuration, with an inspect template that defines what to find and a de-identify template that masks each match, where every infoType in the de-identify template also appears in the inspect template.
Our Professional Agentic Architect practice exams test Model Armor Sensitive Data Protection alongside responsible AI filters and prompt injection controls, with detailed explanations that drill these distinctions.