Security & access
Read-only, by
construction
Zephto cannot change your estate, your workspaces or your model deployments. Not by policy — the grants it holds do not carry write permissions, so there is no configuration mistake that makes it possible.
Access model
What we hold, and for how long
| Control | Position |
|---|---|
| Azure | Two built-in roles: Reader and Cost Management Reader. No custom role definitions, no write or action operations, no Key Vault data plane access. |
| Databricks | Read on the Unity Catalog system tables that carry billing and compute telemetry. We do not read your data tables, notebooks, model artifacts or query results. |
| Fabric | Capacity metrics read. We see CU consumption by workload and workspace — not the contents of your lakehouses, semantic models or reports. |
| Foundry | Token and PTU telemetry through Azure Monitor and Cost Management. We never read prompt or completion content. Token counts, model versions, deployment configuration and cache rates are enough to run every AI check we have. |
| Scope | Set by you. A management group and every workspace, or one subscription and one workspace. We would rather start narrow and earn the rest. |
| Footprint | Nothing deployed into your tenant or workspaces. No agents, no runbooks, no notebooks, no policy assignments. |
| Duration | Every grant is yours to revoke at any point. For one-off assessments we ask you to revoke once the walkthrough call is done. |
Data handling
Where the output lives
Held for comparison, deleted on request
Run output is retained so the next audit can be compared against it, and so we can separate estate change from check-set change. Say the word and it is deleted, including the archived runs.
One tenant’s data stays in one engagement
Findings from your estate are never shown to another client, aggregated into public benchmarks, or used as sales material without your written go-ahead.
We do not read what your models are being asked
Token economics can be audited entirely from counts, configuration and billing telemetry. Prompt and completion bodies are out of scope by design, which keeps the audit clear of your most sensitive AI data.
Reviewing us properly? Ask for the security questionnaire response and the role assignment scripts your team can read before running. Both are available before any access is granted.
Send it to your security team first
We would rather answer the questionnaire up front than halfway through a procurement cycle.