Security controls for Generative AI

Gemini Enterprise Agent Platform implements Google Cloud security controls to help secure your models and training data.

The following table lists the security controls available for Generative AI features:

Data residency (at rest) Customer-managed encryption keys (CMEK) VPC Service Controls (VPC-SC) Access Transparency (AXT)
Gemini 4 Neon
Online prediction
Batch inference
Context caching
Gemini 4 Hydrogen
Online prediction
Batch inference
Context caching
Gemini 4 Carbon
Online prediction
Batch inference
Context caching
Gemini 4 Carbon for cyber
Online prediction
Batch inference
Context caching
Gemini 4 Argon
Online prediction
Batch inference
Context caching
Gemini 4 Argon Cyber Permissive
Online prediction
Batch inference
Context caching
Gemini 3.8 Live
Online prediction
Gemini 3.8 Flash Cyber
Online prediction
Context caching
Gemini 3.8 Flash
Online prediction
Batch inference
Context caching
Gemini Omni 1.2 Flash
Online prediction
Gemini 3.7 Flash
Online prediction
Batch inference
Context caching
Gemini 3.6 Flash
Online prediction
Batch inference
Context caching
Gemini 3.5 Flash-Lite
Online prediction
Batch inference
Context caching
Gemini 3.1 Flash-Lite Image (Nano Banana 2 Lite)
Online prediction
Batch inference
Context caching
Gemini 3 Pro Image
Online prediction
Batch inference
Context caching
Gemini Nano Banana 2.1
Online prediction
Batch inference
Context caching
Gemini 3.1 Flash Image
Online prediction
Batch inference
Context caching
Gemini 3.5 Flash
Online prediction
Batch inference
Tuning
Context caching
Gemini 3.1 Flash-Lite
Online prediction
Batch inference
Context caching
Gemini 2.5 Pro
Online prediction
Batch inference
Tuning
Context caching
RAG Engine
Grounding with Google Search and Grounding with Google Maps
Gemini 2.5 Flash Image
Online prediction
Batch inference
Gemini 2.5 Flash
Online prediction
Batch inference
Tuning
Context caching
RAG Engine
Grounding with Google Search and Grounding with Google Maps
Gemini 2.5 Flash-Lite
Online prediction
Batch inference
Tuning
Context caching
RAG Engine
Grounding with Google Search and Grounding with Google Maps
Gemini 2.5 Flash with Gemini Live API native audio
Online prediction
Gemini 3.5 Transcribe
Online prediction
Deferred tier
Deferred model calls
Gemini Embedding 2
Online prediction
Embeddings for Multimodal
Online prediction
Batch prediction
Tuning
Embeddings for Text
Online prediction
Batch prediction
Tuning
RAG Engine
Veo for Video Generation
Online prediction
Partner Models (MaaS)
Online prediction
Open Models (MaaS)
Online prediction

Prompt logging for Partner and Open Models (MaaS) is turned off by default and adheres to the Agent Platform Zero Data Retention policy.



Deferred tier controls

The Deferred tier queues turns and persists their state on the global endpoint. In addition to the controls in the preceding table, the following controls apply to deferred model calls:

  • Supported: Cloud Audit Logs and FedRAMP High.
  • VPC-SC restriction: External mcp_server tools aren't allowed inside a VPC-SC perimeter.
  • Not supported: Stateless zero data retention (ZDR), DoD IL5, and ITAR. Deferred execution requires store=true, so setting store=false or enforcing the constraints/vertexai.disableAPIStorage organization policy fails with HTTP 400 ("code": "invalid_request").