
Model Armor screens the prompts going into a generative AI model and the responses coming out. On the Professional Agentic Architect exam, it is worth being precise about what content screening covers and where it fits in an agent architecture. Model Armor evaluates content inside prompts, responses, and specific MCP requests, rather than controlling data movement between projects, authorizing which tools an agent can call, or scanning application source code. Architectural decisions around agent security often come down to configuring Model Armor templates properly and separating content screening from adjacent Google Cloud controls.
Model Armor organizes its filters inside regional templates. A template can combine four categories of content checks:
Each template sets an enforcement type. Inspect and block stops the whole violating response, while inspect only detects violations without blocking anything. Confidence levels apply only to the responsible AI and prompt injection filters.
Within a Model Armor template, Sensitive Data Protection runs in either a basic or an advanced configuration. Basic Sensitive Data Protection supports inspection only, so it reports sensitive data such as card numbers without changing the content that is returned.
Advanced Sensitive Data Protection adds the ability to mask. It references both an inspect template and a de-identify template. The de-identify template masks each match while letting the remaining text go through, which keeps the useful part of a response that blocking would discard. Every infoType in the de-identify template must also appear in the inspect template.
Model Armor can be applied in two broad ways, as a policy attached to Agent Gateway or as an API that the agent's own code calls. Agent Gateway has two modes, and each one sees different traffic. A Client-to-Agent gateway sits in front of an agent and handles the traffic that users send to it. An Agent-to-Anywhere gateway handles the agent's outbound traffic to tools and other agents. A Model Armor template attached to a gateway screens the traffic that passes through that gateway, so the gateway's mode decides whether user prompts or outbound calls are inspected.
A few configuration rules apply to gateway attachment. The template must be in the same region as the gateway. Agent Gateway works with Model Armor templates rather than floor settings, which play a different role by defining the minimum configuration that templates must meet. On a gateway handling MCP traffic, Model Armor screens tools/call and prompts/get requests, and it does not screen tools/list. Because the policy lives in the gateway and the template, it is managed separately from the agent's code.
The alternative is to call the sanitize API from a callback in the agent's code. In that pattern Model Armor acts as a detector only. It returns its findings, the agent's own code decides what to block, and the policy is maintained in the codebase along with the agent.
On the Professional Agentic Architect exam, Model Armor sits alongside several Google Cloud controls that solve different security problems. Because Model Armor screens content in prompts, responses, and screened MCP calls, it does not replace controls built for project perimeters, tool authorization, or repository security.
Model Armor has no control over data moving between projects through Google Cloud APIs. That is the job of VPC Service Controls. A perimeter restricts Google Cloud APIs, such as BigQuery and Cloud Storage, at the boundary of a group of projects, and it applies even when the caller holds valid IAM permissions. Controls also differ in what they are scoped to. A perimeter is scoped to projects and APIs, so it applies to every caller. Identity-scoped controls, such as a principal access boundary policy or an Agent Gateway binding, apply only to the principals or agents they are attached to.
Model Armor also does not decide which servers or tools an identity may invoke. Outbound tool access belongs to Agent Registry and Agent Gateway. An Agent-to-Anywhere gateway checks every destination against Agent Registry and evaluates egress access policies for the agent's Agent Identity, denying requests by default when a server is not registered or no matching policy exists. Because Agent Gateway parses MCP traffic, an egress access policy can target individual tools rather than whole servers. Policy conditions can test tool attributes such as iap.googleapis.com/mcp.tool.isReadOnly, which matches tools marked as read-only. A policy enforced at the gateway sits outside the agent, so it holds regardless of later changes to the agent's code, which is the key difference from narrowing an agent's tools inside its own McpToolset configuration.
Finally, Model Armor does not scan application source code for vulnerabilities. That work belongs to CodeMender in Agent Platform (in Public Preview for a limited set of customers), an AI code security agent that finds, verifies, and fixes vulnerabilities in a repository, with Java among its supported languages. It verifies a finding by building the code and attempting an exploit, and it generates and tests a patch. Human confirmation of its actions is on by default, so engineers review changes before they are applied.
Our Professional Agentic Architect practice exams test Model Armor alongside Agent Gateway access policies and VPC Service Controls, with detailed explanations that drill these distinctions.