Google Cloud Model Armor Features

Provides specific capability to defend AI assets from adversarial attacks (including prompt injection, data poisoning, model inversion or extraction) without requiring retraining of the underlying model.

Enables automatic inspection of model inputs (prompts) and/or outputs (responses) to prevent unsafe, sensitive or manipulated content.

Monitors the integrity of model weights, dependencies or metadata (for example via SBOM/AIBOM) to detect tampering, drift or unauthorised modification.

Verifies the ability to enforce who or what (users, agents, systems) may access a model or LLM asset.

Detects unusual or harmful runtime behaviour of AI models, agents or workflows (such as unexpected output patterns, excessive permissions use, or unknown agents).

Provides a persistent, searchable audit log of AI‑asset inputs, outputs and interactions (including who/what invoked the model, when, and with what data) for forensics and compliance.

Ensures that the AI‑security platform supports scaling of AI‑asset protection (models, agents, multi‑cloud deployments) and applies governance/compliance frameworks as AI usage grows.

Enables integration of the AI‑security solution with traditional security stacks (SIEM, SOAR, cloud security, application security, identity/access management) to unify visibility and response.

Offers visibility into unmanaged or unauthorized AI/agent use (“shadow AI”) across the organisation and enforces control over such usage (e.g., agent creation, LLM‑based services).

Supports codified, machine‑enforceable security policies targeting AI models/agents (for example, blocking certain categories of prompts, enforcing least‑privilege for model use, enforcing “no external data” rules).