One security baseline across Azure, GCP and Alibaba Cloud
Multi-cloud does not mean three security programmes. It means one set of control objectives, each with a per-provider implementation and a per-provider evidence query.
By Piotr Zenon Żabrowski · Published · Updated
1Governance
- One control set
- Framework mapping
- Exception register
2Identity
- Central IdP
- Federation
- No long-lived keys
3Platform
- Landing zones
- Policy as code
- Network isolation
4Detection
- Central SIEM
- Per-cloud log sources
- ATT&CK coverage
Objective-first, provider-second
| Control objective | Azure | Google Cloud | Alibaba Cloud |
|---|---|---|---|
| Central identity, no local accounts | Entra ID with Conditional Access | Cloud Identity with context-aware access | RAM users federated to the corporate IdP |
| Time-bound privilege | PIM eligible role assignments | IAM conditions plus short-lived grants | RAM policy with STS-issued temporary credentials |
| Guardrails inherited by every workload | Azure Policy at management-group scope | Organization policy constraints | Resource Directory policies |
| Complete audit trail off the workload account | Activity log to a Sentinel workspace | Cloud Audit Logs to a log sink project | ActionTrail to a dedicated logging account |
| No default network exposure | NSG deny-by-default plus Private Endpoints | VPC-SC plus Private Service Connect | Security group deny-by-default plus PrivateLink |
| Encryption with managed keys | Key Vault with rotation policy | Cloud KMS with rotation | KMS with automatic rotation |
| Runtime guardrails for AI workloads | Prompt Shields via Azure AI Content Safety | Model Armor for prompt and response screening | AI Guardrails content and prompt moderation |
What actually differs
- Terminology: the same concept has three names, which is why the objective layer must be provider-neutral.
- Default posture: providers differ on what is public or permissive out of the box — verify, never assume.
- Regional and data-residency behaviour, which matters most for Alibaba Cloud deployments serving EU users.
- Detection coverage: native detections vary in depth, so normalise into one SIEM rather than living in three consoles.
- Guardrail maturity for AI-specific controls: the underlying moderation and injection-detection models differ across providers even when the objective and integration point are identical, so validate detection quality per provider rather than assuming feature parity.
Frequently asked questions
Is one SIEM realistic across three clouds?
Yes, and it is the main thing that makes multi-cloud operable. Normalise identity and audit events first; resource telemetry can follow.
Which framework to anchor on?
CIS Benchmarks per provider for technical detail, NIST CSF or ISO 27001 for the objective layer that management reports against.
Where do AI workloads fit in this baseline?
Yes, and it belongs in the same objective-first model rather than as a separate programme. Each provider now ships a native runtime guardrail for generative and agentic AI workloads — Model Armor on Google Cloud, Prompt Shields within Azure AI Content Safety, and AI Guardrails on Alibaba Cloud — so the objective 'screen prompts and model output at runtime' gets a provider-specific answer just like every other row in the table below.
Related
Cloud security
Multi-cloud security practice: landing zones, identity, posture management, MITRE ATT&CK-aligned detection engineering and CIS hardening baselines.
Microsoft Azure
Azure landing zone design, Cloud Adoption Framework alignment, policy rollout sequencing and management group governance.
Google Cloud
Google Cloud IAM hardening: organization policy design, workload identity federation, VPC Service Controls perimeters and agent identity.
Alibaba Cloud
Alibaba Cloud security review: RAM permission audit, AccessKey lifecycle, Security Center configuration and cross-border data flow compliance.
Cloud architecture review
Design-stage review of cloud and hybrid architecture: trust boundaries, identity model, segmentation, key management and detection design.
ISSA Polska Wrocław
Five talks given at the Wrocław chapter of ISSA Polska between 2022 and 2025: OSINT, threat modeling production LLMs, CloudSec and device code phishing.