What Model Armor Actually Does
Google Cloud documents Model Armor as a service that screens LLM prompts and responses to help protect against prompt injection, jailbreak attempts, sensitive data exposure, and malicious content. It integrates with Vertex AI and can be applied to traffic passing through Apigee API management or Google Cloud network security products. In practical terms, Model Armor sits at the model input and output boundary. It inspects text before it reaches a model and after a model produces a response, applying classifier-based detection to flag or block content that matches known risk patterns.
This is a content-safety and data-loss-prevention function. It is positioned in Google Cloud's documentation alongside Vertex AI safety filters and Security Command Center, which reinforces that its intended role is inspection of text content rather than control over what an agent does with that content once processing continues. Google's published guidance also distinguishes Model Armor's filtering role from IAM-based access control, indicating that identity and permission management remain the responsibility of separate Google Cloud IAM services.
What Falls Outside Model Armor's Scope
Model Armor's documented scope does not extend to agent identity management, least-privilege credential issuance, or policy enforcement on individual tool or function calls. This matters because modern AI agents rarely stop at generating text. Many use frameworks such as the Model Context Protocol (MCP) to invoke external tools, query APIs, or take actions on behalf of a user. Those tool-call interactions represent a separate execution path that is distinct from the prompt and response text Model Armor inspects.
No official Google Cloud statement was found explicitly addressing how Model Armor interacts with MCP-based tool calling. Based on its documented architecture, Model Armor's inspection point is the text boundary, not the tool-invocation layer, so agents that pass content screening can still execute actions that a content filter has no visibility into or authority over.
Where Trussed Operates in the Stack
Trussed provides runtime governance for AI agents, addressing the layer where agents take action rather than the layer where they generate or receive text. This includes agent identity, permission scoping, least-privilege enforcement, and tool approval workflows for MCP and other tool-calling frameworks. Runtime policy enforcement at this layer determines whether a given agent, in a given session, is authorized to invoke a specific tool or perform a specific action, independent of whether the surrounding prompt or response text passed a content filter.
Trussed also provides audit logging at the level of agent actions and tool calls, which supports investigation and compliance needs that prompt and response logging alone does not satisfy. This is a different function from content screening: it does not evaluate whether text contains a jailbreak attempt, it evaluates whether an authenticated agent identity has the permission to execute a requested action.
Enforcement Point Comparison
| Capability | Model Armor | Trussed |
|---|---|---|
| Inspection point | Prompt and response text (input/output boundary) | Agent identity and tool-call execution (runtime) |
| Primary risk addressed | Injection, jailbreak, sensitive data exposure | Unauthorized actions, excess permission scope |
| Coverage path | Traffic routed through Vertex AI or Apigee | Any agent invoking tools or MCP functions |
| Audit granularity | Prompt/response logging | Action-level and tool-call logging |
| Identity and permissions | Not addressed (handled by separate IAM services) | Core function: least-privilege enforcement |
Complementary, Not Redundant
Based on documented capabilities, Model Armor and Trussed address different risk categories and are architecturally complementary rather than overlapping. Model Armor reduces the likelihood that malicious or unsafe text reaches a model or a user. Trussed reduces the likelihood that an agent, once operating, takes an unauthorized action or accesses a resource beyond its granted scope. An architecture that relies on only one of these layers leaves a gap: content filtering without runtime governance permits an agent that passed inspection to still act outside its intended permissions, while runtime governance without content filtering leaves prompt-level attacks such as injection or jailbreak attempts unaddressed before they influence model behavior.
Organizations building on Google Cloud with Vertex AI or Apigee-routed traffic can use Model Armor for its documented purpose. Agents that call tools, invoke MCP-based functions, or operate across multiple cloud or model backends need a runtime governance layer to enforce identity and permission boundaries that Model Armor does not address.