Cloud platforms
Secure AI deployment on every major cloud
Your AI is only as safe as the cloud configuration underneath it. We publish the configuration we expect on each platform — and certify that yours matches it.
Microsoft Azure
Azure OpenAI in Azure AI Foundry, Azure AI Search, Content Safety, API Management and Azure Machine Learning.
View Azure baseline AWSAmazon Web Services
Amazon Bedrock, Bedrock Guardrails, Knowledge Bases and Agents, SageMaker AI and the AWS security services around them.
View AWS baseline GCGoogle Cloud
Vertex AI, Gemini models, Model Armor, Vertex AI Search and the Google Cloud controls that contain them.
View Google Cloud baseline +Other platforms
Oracle OCI, Databricks, Snowflake, OpenAI and Anthropic APIs, self-hosted models and Microsoft 365 Copilot.
View other platformsShared responsibility
The cloud secures the model. You secure the deployment.
Managed AI services like Azure OpenAI, Amazon Bedrock and Vertex AI are secure by design — but almost every setting that decides whether your deployment is safe is yours to configure.
Cloud provider
- Physical & host security
- Model hosting and isolation
- Service availability
- Not training on your prompts (per provider terms)
You — and what we certify
- Who and what can call your models
- Public vs. private endpoints
- Encryption keys and data retention
- Guardrails, quotas and gateways
- Data your AI can retrieve
- Agent and tool permissions
- Logging, alerting and response
Cross-cloud comparison
The same controls, in each cloud's language
Every platform can meet the CertiSafe AI baseline. Here's what each control looks like on Azure, AWS and Google Cloud.
| Control | Microsoft Azure | AWS | Google Cloud |
|---|---|---|---|
| Managed model service | Azure OpenAI in Azure AI Foundry | Amazon Bedrock | Vertex AI |
| Keyless identity | Microsoft Entra ID + managed identities; local (key) auth disabled | IAM roles with temporary credentials; no long-term Bedrock API keys | Service accounts + Workload Identity Federation; service account keys blocked |
| Least-privilege access | Scoped Azure RBAC roles (e.g. Cognitive Services OpenAI User) | IAM policies scoped to approved model ARNs; SCPs deny the rest | Narrow Vertex AI roles; Organization Policy restricts allowed models |
| Private networking | Private endpoints; public network access disabled | VPC interface endpoints (AWS PrivateLink) with endpoint policies | VPC Service Controls perimeter; Private Service Connect |
| Customer-managed keys | Azure Key Vault (CMK) | AWS KMS customer managed keys | Cloud KMS (CMEK) |
| Guardrails | Azure AI Content Safety filters + Prompt Shields | Amazon Bedrock Guardrails, enforced via IAM | Model Armor + Gemini safety settings |
| AI gateway & quotas | Azure API Management AI gateway (token limits) | API gateway / proxy tier + service quotas | Apigee or API gateway + quotas |
| Sensitive data protection | Microsoft Purview (DSPM for AI) | Guardrails sensitive-info filters; Amazon Macie for S3 | Sensitive Data Protection (DLP) |
| Audit logging | Diagnostic settings → Log Analytics; Activity Log | Bedrock model invocation logging + CloudTrail | Cloud Audit Logs (incl. Data Access logs) |
| Threat detection | Microsoft Defender for Cloud — AI threat protection | Amazon GuardDuty + AWS Security Hub | Security Command Center |
| Policy as code | Azure Policy | Service control policies + AWS Config rules | Organization Policy Service |
| RAG / retrieval store | Azure AI Search with security filters | Bedrock Knowledge Bases with metadata filtering | Vertex AI Search / RAG Engine with access controls |
Cloud services evolve quickly. Our audit baselines are reviewed every quarter against each provider's current security guidance.
How we certify a cloud deployment
Configuration review and attack testing, together
A configuration scan alone misses how your AI actually behaves. Attack testing alone misses the misconfigurations waiting to be found. We do both.
Read-only access
You grant a time-boxed, read-only identity to the in-scope accounts, subscriptions or projects. No write access, no customer data.
Configuration review
Automated and manual review of every AI resource against the platform baseline: identity, network, keys, guardrails, logging.
Attack testing
Adversarial testing of the AI applications running on that infrastructure — injection, data leakage, agent abuse.
Evidence & certificate
Findings, remediation re-checks, then a certificate and evidence pack you can hand to auditors and customers.
Frequently asked questions
Which cloud platforms does CertiSafe AI certify?
We certify AI deployments on Microsoft Azure (Azure OpenAI in Azure AI Foundry, Azure AI Search, Azure Machine Learning), Amazon Web Services (Amazon Bedrock, SageMaker), and Google Cloud (Vertex AI), as well as Oracle Cloud Infrastructure, Databricks, Snowflake, direct model APIs such as OpenAI and Anthropic, self-hosted open-weight models, and Microsoft 365 Copilot. Multi-cloud environments are certified as a single scope.
Who is responsible for AI security in the cloud — us or the cloud provider?
Both, under the shared responsibility model. The provider secures the underlying infrastructure and the managed model service. You are responsible for how you configure it: who can call the models, whether endpoints are public, which data the AI can reach, which guardrails are enforced, and whether activity is logged. Almost every real-world AI cloud incident is a customer-side configuration issue.
Do you need write access to our cloud accounts?
No. We use a time-boxed, read-only identity scoped to the accounts, subscriptions or projects in scope. We review configuration and metadata, not the content of your prompts or customer data, and access is removed when the audit ends.
Can a multi-cloud AI deployment be certified?
Yes. Many organizations use Azure OpenAI for one workload and Amazon Bedrock or Vertex AI for another. We audit each platform against its own baseline and issue one certification covering the whole in-scope AI footprint.
Running AI on more than one cloud?
We'll scope your whole AI footprint — every cloud, every model, every app — into one certification.