Security
Controls you can verify. Limits we state plainly.
Synthetic data is a confidential asset. Automated checks reduce risk; they cannot prove that nothing about the original can be learned, and a downloaded file cannot be recalled.
Synth runs inside your cloud account, under your keys and your single sign-on.
What this build does
| Capability | Status | Notes |
|---|---|---|
| Deployment in your own cloud or data centre, in the region you choose | live | Container image, Helm chart and Terraform for AWS, Azure and Google Cloud, or Docker Compose on one host. A one-step installer for Azure. Runs as a single replica: no high availability yet. |
| Single sign-on | live | OpenID Connect with Microsoft Entra ID, Okta, Google Workspace and others; roles follow your groups; can be required for everyone except named break-glass administrators. SAML and SCIM provisioning are not included. |
| Managed keys | live | Every key is wrapped by your key service (AWS KMS, Azure Key Vault HSM keys, Google Cloud KMS or HashiCorp Vault). A separate key for real data acts as a kill switch. |
| Assistant (built-in, rules-based) | live | Runs inside the deployment; no data leaves. |
| Assistant (Claude via Anthropic API) | available | Optional. Sees descriptions and synthetic aggregates only; never records. Enabled per organisation by an administrator. |
| File upload and schema-only definitions | live | Uploads are encrypted on receipt; the web and assistant processes cannot read them back. |
| Watched sources: landing folders and database files | live | Folders of CSV exports and read-only SQLite database files, mounted read-only. The coverage agent keeps a synthetic version of every table current; every version still needs the owner’s approval. |
| PostgreSQL, object storage and data warehouse connectors | not available | Defined but not built. Export to a watched landing folder in the meantime. |
| Synthetic generation with quality, privacy and usefulness checks | live | Three methods, including a differentially private option with a privacy budget. Checks are empirical indicators, not guarantees. |
| Partner workspaces with approved analyses | live | Restricted worker on synthetic data only; budgets and output limits. |
| Real-data model testing | live | Separate service; partners see reviewed, rounded results only. |
| Straight-through testing of your own applications | live | Fresh synthetic test packs with edge cases; your CI runs the application and Synth scores the results. |
| Test databases and MCP for partners’ coding agents | live | Pinned synthetic versions, approved test suites and registered builds; every result is reviewed before the partner sees it. Direct read-only database access only where you allow it. |
| Partner application containers in a sandbox | available | One isolated virtual machine per test run, with no network beyond the platform. Off until the sandbox node pool is installed and your administrator records its validation. |
| Signed hand-over into your systems | live | Reviewed partner deliverables, signed and collected by your integration gateway. Synth never runs them on your real data. |
| Partner code execution in analysis workspaces | disabled | No isolated runtime is configured. Vendor code execution is disabled (fails closed). |