Google Cloud
Secure AI deployment on Google Cloud
The configuration we expect for Vertex AI, Gemini and the Google Cloud services around them — and exactly how we verify it before issuing a certificate.
Quick answer
A secure Google Cloud AI deployment blocks service account key creation and gives each workload a dedicated least-privilege service account, places Vertex AI and its data inside a VPC Service Controls perimeter, encrypts AI resources with CMEK, screens prompts and responses with Model Armor, enforces document-level access in retrieval, and enables Data Access audit logs for Vertex AI with Security Command Center watching for threats.
Services in scope
- Vertex AI
- Gemini on Vertex AI
- Model Garden
- Model Armor
- Vertex AI Search
- Vertex AI RAG Engine
- Vertex AI Agent Engine
- Vertex AI Workbench
- Cloud Storage & BigQuery (AI data)
- Cloud KMS
- Security Command Center
Legend: Required must pass for certification. Assessed is tested and reported with recommendations, and becomes required where your data classification demands it.
Identity & access
| Control | Recommended configuration | How we verify |
|---|---|---|
| No service account keys Required | Organization Policy iam.disableServiceAccountKeyCreation enforced. Workloads use attached service accounts; external workloads use Workload Identity Federation. | Org policy review; inventory of existing user-managed keys. |
| Least-privilege service accounts Required | A dedicated service account per AI workload with narrow Vertex AI roles such as roles/aiplatform.user. No basic Owner or Editor roles, and no use of the default compute service account. | IAM policy analysis across organization, folders and projects. |
| Privileged access Required | Admin roles granted to groups, time-bound through Privileged Access Manager or IAM Conditions, with 2-Step Verification enforced. | IAM and identity configuration review. |
| Approved models only Assessed | Organization Policy restricts which Model Garden models can be used. | Org policy review. |
Network isolation
| Control | Recommended configuration | How we verify |
|---|---|---|
| VPC Service Controls Required | A service perimeter around AI projects that restricts aiplatform.googleapis.com, storage.googleapis.com and bigquery.googleapis.com, with explicit ingress and egress rules. | Perimeter configuration and dry-run violation review. |
| Private connectivity Assessed | Workloads reach Google APIs over Private Google Access or Private Service Connect. Online prediction uses private endpoints where available. | Network and endpoint configuration review. |
| Workbench hardening Assessed | Org policies ainotebooks.restrictPublicIp and ainotebooks.disableRootAccess enforced on Vertex AI Workbench. | Org policy and instance review. |
Data protection
| Control | Recommended configuration | How we verify |
|---|---|---|
| Data residency Required | Organization Policy gcp.resourceLocations set. Regional vs. global model endpoints chosen deliberately to match where data may be processed. | Org policy and endpoint usage review. |
| CMEK Assessed | Cloud KMS customer-managed keys for Vertex AI datasets, endpoints, pipelines and RAG corpora, enforced with gcp.restrictNonCmekServices. Required for regulated data. | Resource encryption and org policy review. |
| Retention decisions Assessed | Documented decision on Gemini data caching, abuse-monitoring logging and zero-data-retention configuration. | Documentation and project configuration review. |
| De-identification Assessed | Sensitive Data Protection de-identifies data in ingestion pipelines feeding RAG and tuning. | Pipeline review. |
Guardrails
| Control | Recommended configuration | How we verify |
|---|---|---|
| Model Armor Required | Model Armor templates screen prompts and responses for prompt injection and jailbreaks, sensitive data, malicious URLs and harmful content. | Template and integration review; adversarial testing. |
| Safety settings Required | Gemini safety settings are not lowered to BLOCK_NONE or OFF without documented approval. | Code and configuration review. |
| Grounding Assessed | Customer-facing answers are grounded in approved sources, with citations. | Output testing. |
Rate limits & cost
| Control | Recommended configuration | How we verify |
|---|---|---|
| Per-consumer limits Required | Per-user or per-tenant quotas at the API tier (Apigee, API Gateway or an AI gateway). | Configuration review and controlled load testing. |
| Quotas & budgets Assessed | Vertex AI quotas sized deliberately; Cloud Billing budgets with alerts on AI projects. | Quota and budget review. |
Retrieval (RAG)
| Control | Recommended configuration | How we verify |
|---|---|---|
| Document-level access Required | Vertex AI Search data store access controls or server-side filters derived from the user's identity limit results to what the user may see. | Configuration and code review; cross-user retrieval testing. |
| Separated corpora Required | Public, internal and per-tenant content live in separate data stores or RAG corpora. | Data store inventory. |
Agents & tools
| Control | Recommended configuration | How we verify |
|---|---|---|
| Least-privilege agents Required | Agents deployed on Agent Engine or Cloud Run use dedicated service accounts with only the permissions their tools need. | Service account and tool review. |
| Human approval Required | Sensitive actions require user confirmation and server-side authorization. | Flow review; agent abuse testing. |
Logging & threat detection
| Control | Recommended configuration | How we verify |
|---|---|---|
| Data Access audit logs Required | DATA_READ and DATA_WRITE audit logs enabled for aiplatform.googleapis.com. | Audit configuration review at org, folder and project level. |
| Central log retention Assessed | Aggregated log sinks to a central project, with locked retention on the log bucket. | Sink and bucket review. |
| Threat detection Required | Security Command Center enabled for the organization, with findings routed to your responders. | SCC tier, services and notification review. |
Reference configuration
Block service account keys across the organization, and put AI projects inside a VPC Service Controls perimeter:
gcloud resource-manager org-policies enable-enforce \
iam.disableServiceAccountKeyCreation --organization=ORGANIZATION_ID
gcloud access-context-manager perimeters create ai_perimeter \
--title="AI workloads" \
--resources=projects/PROJECT_NUMBER \
--restricted-services=aiplatform.googleapis.com,storage.googleapis.com,bigquery.googleapis.com \
--policy=ACCESS_POLICY_IDEnable Data Access audit logs for Vertex AI in your IAM policy's audit configuration:
auditConfigs: - service: aiplatform.googleapis.com auditLogConfigs: - logType: ADMIN_READ - logType: DATA_READ - logType: DATA_WRITE
Check yourself: find user-managed service account keys in a project:
gcloud iam service-accounts list --format="value(email)" | while read sa; do gcloud iam service-accounts keys list --iam-account="$sa" --managed-by=user done
How we certify a Google Cloud deployment
- Read-only accessYou grant a group or service account the Security Reviewer and Cloud Asset Viewer roles on the organization or folder containing your AI projects.
- Full AI inventoryCloud Asset Inventory discovery of every Vertex AI endpoint, model, data store, notebook and AI data bucket across all in-scope projects.
- Configuration auditEvery resource, org policy and perimeter is checked against this baseline, cross-referenced with Security Command Center findings.
- Attack testingAdversarial testing of the applications and agents running on the infrastructure: prompt injection, data leakage, retrieval controls and tool abuse.
- Remediate, verify, certifyYou receive findings with exact settings to change. We re-verify, issue your certificate and evidence pack, and re-audit every quarter.
Frequently asked questions
What is the most important Vertex AI security control?
A VPC Service Controls perimeter around the projects that use Vertex AI, with aiplatform.googleapis.com and your data services (Cloud Storage, BigQuery) as restricted services. Even if credentials leak, data can't be pulled out to resources outside the perimeter.
Are Vertex AI Data Access audit logs on by default?
No. Admin Activity audit logs are always on, but Data Access audit logs — which record calls such as model predictions — must be enabled for aiplatform.googleapis.com in your project, folder or organization audit configuration.
What is Model Armor?
Model Armor is a Google Cloud service that screens prompts and responses for prompt injection and jailbreak attempts, sensitive data, malicious URLs and harmful content, using templates you define. It can be applied to Vertex AI and to other models your applications call.
What access does CertiSafe AI need to certify our Google Cloud environment?
A Google group or service account granted the Security Reviewer and Cloud Asset Viewer roles at the organization or folder level that contains your AI projects. It's read-only and removed when the audit ends.
Certify your Google Cloud AI deployment
Most Google Cloud environments can be audited within two weeks of a scoping call.
Other platforms
Compare all platformsMicrosoft Azure
Azure OpenAI, AI Foundry, AI Search, Content Safety and API Management.
View baseline AWSAmazon Web Services
Amazon Bedrock, Guardrails, Knowledge Bases, Agents and SageMaker AI.
View baseline +Other platforms
OCI, Databricks, Snowflake, model APIs, self-hosted models and Copilot.
View baseline