Other platforms
AI runs in more places than your cloud console
Data platforms, direct model APIs, self-hosted models and workplace copilots all belong in scope. Here's what we expect on each — and what we check.
Quick answer
Whatever the platform, a secure AI deployment needs the same things: keyless or tightly scoped identity, private network paths, data governance that carries through to the model, guardrails and rate limits, retrieval that respects user permissions, and audit logs. On data platforms that means using their native governance (Unity Catalog, Snowflake roles). For direct model APIs it means server-side, project-scoped keys. For self-hosted models it means an authenticated gateway in front of every inference server.
Oracle Cloud Infrastructure (OCI Generative AI)
| Control | Recommended configuration | How we verify |
|---|---|---|
| Compartment-scoped IAM Required | Generative AI permissions granted to specific groups or dynamic groups in specific compartments, for example allow group ai-apps to use generative-ai-family in compartment ai-prod. No tenancy-wide grants. | IAM policy review. |
| Private access Required | Generative AI accessed through private endpoints or a service gateway, not over the public internet. | Network configuration review. |
| Keys & detection Assessed | OCI Vault for keys and secrets. OCI Audit retained. Cloud Guard enabled with responders. | Service configuration review. |
Databricks (Mosaic AI)
| Control | Recommended configuration | How we verify |
|---|---|---|
| Unity Catalog governance Required | Models, feature tables and vector search indexes are governed in Unity Catalog with least-privilege grants. Row filters and column masks apply to data used for RAG. | Catalog grant and policy review. |
| AI Gateway Required | Serving endpoints use Mosaic AI Gateway rate limits, guardrails and usage tracking, with inference tables capturing requests to a governed location. | Endpoint and gateway configuration review. |
| Identity & network Required | Service principals with OAuth instead of personal access tokens. Private Link and IP access lists on workspaces. | Workspace security configuration review. |
Snowflake (Cortex AI)
| Control | Recommended configuration | How we verify |
|---|---|---|
| Restricted Cortex access Required | The SNOWFLAKE.CORTEX_USER database role is revoked from PUBLIC and granted only to roles that need AI functions. | Grant review. |
| Data policies apply Required | Masking and row access policies protect sensitive columns before data reaches Cortex functions or Cortex Search. | Policy coverage review; test queries. |
| Cross-region inference Assessed | The CORTEX_ENABLED_CROSS_REGION account parameter is set deliberately to match data residency requirements. | Account parameter review. |
OpenAI, Anthropic & other model APIs
| Control | Recommended configuration | How we verify |
|---|---|---|
| Server-side, scoped keys Required | Separate project- or workspace-scoped keys per application and environment, stored in a secrets manager and used only from your servers — never in browsers, mobile apps or client code. | Key inventory via the provider's admin console; client bundle and repository scanning. |
| Console access Required | Provider organization accounts use SSO with MFA. Admin roles are limited, and departed staff are removed. | Organization member and role review. |
| Spend & rate limits Required | Spend limits per project, plus per-user rate limits in your own application tier. | Console and application configuration review. |
| Data retention terms Assessed | Data retention and training settings match your obligations, with zero-data-retention arrangements where eligible and required. | Contract and settings review. |
| Leak detection Assessed | Secret scanning enabled on code repositories, with a documented key-revocation runbook. | Repository settings review. |
Self-hosted & open-weight models
Running Llama, Mistral, Qwen or other open-weight models on your own GPUs — with vLLM, Ollama, Text Generation Inference or similar — puts every security responsibility on you.
| Control | Recommended configuration | How we verify |
|---|---|---|
| No exposed inference servers Required | Inference servers are never reachable from the internet. Many ship with no authentication — Ollama's API on port 11434, for example. Access goes through an authenticating gateway. | External port scanning and network policy review. |
| Model provenance Required | Weights come from verified sources with recorded hashes, in safe formats such as safetensors. No loading of untrusted pickle-based checkpoints. | Model registry and artifact review. |
| Workload isolation Required | Inference runs in dedicated namespaces or nodes, with Kubernetes network policies, non-root containers and scanned images. | Cluster configuration review. |
| Guardrails & patching Assessed | An input/output guardrail layer in front of the model. The serving stack is patched on a defined cadence. | Architecture review and version inventory. |
Workplace copilots (Microsoft 365 Copilot, Gemini for Workspace)
| Control | Recommended configuration | How we verify |
|---|---|---|
| Oversharing remediation Required | Broadly shared sites, drives and "anyone with the link" files are reviewed and fixed before rollout. Restricted Content Discovery or similar controls are used where cleanup isn't finished. | Sharing and permission reports; sampled Copilot queries with test users. |
| Labels & DLP Required | Sensitivity labels and data loss prevention policies (Microsoft Purview, Google Workspace DLP) cover confidential content. | Label and policy coverage review. |
| Audit Assessed | Copilot interaction audit logs are retained and reviewable. | Audit configuration review. |
How we certify these platforms
- Read-only accessA read-only or auditor role on each platform: OCI auditor policies, Databricks workspace and account read access, a Snowflake role with read-only access to account usage, admin-console read access for model APIs, and cluster read access for self-hosted models.
- InventoryWe discover every AI resource, key, endpoint and model — including the ones nobody listed.
- Configuration auditEach platform is checked against the controls above.
- Attack testingThe applications and agents built on these platforms get the same adversarial testing as any cloud deployment.
- One certificateAll platforms in scope roll up into a single certification and evidence pack, re-audited every quarter.
Frequently asked questions
Is it safe to run open-weight models like Llama or Mistral ourselves?
It can be, but self-hosting moves every security responsibility to you. Most inference servers ship without authentication. Never expose them directly to the internet, put them behind an authenticating gateway, load model weights only from verified sources in safe formats such as safetensors, and patch the serving stack like any other production software.
How should we secure OpenAI or Anthropic API keys?
Use separate, project-scoped keys per application and environment, store them in a secrets manager, call the API only from your servers (never from browsers or mobile apps), set spend limits, enforce SSO and MFA on the provider console, and enable secret scanning in your code repositories so leaked keys are caught and revoked quickly.
Why is Microsoft 365 Copilot a security concern?
Copilot can surface any content the signed-in user already has permission to access. In many organizations, years of overshared SharePoint sites, Teams and OneDrive files mean users technically have access to far more than they should. Copilot makes that oversharing easy to find, so permissions and sensitivity labels need cleaning up before and during rollout.
Running AI somewhere not listed here?
If it touches your data or your customers, we can scope it. Tell us what you run.