Pre-release software
DataNivra is pre-release software. Its privacy boundary — zero raw-production-data egress through customer-resident processing — is implemented and tested in the product's own security suite, but it has not been assessed by an independent auditor. DataNivra holds no certification and no SOC 2 report or ISO 27001 certificate. Industry packs support your own HIPAA, GDPR or PCI DSS work; using them does not make an organisation compliant, and masking alone is not proof of anonymization.
Product limitations
- Connectors. Only the connectors marked Available now on Integrations ship in the agent. The others are planned extension points; asking for one fails with
CONNECTOR_NOT_AVAILABLE. Read-only access is proven automatically for PostgreSQL and file systems; object stores and other SQL dialects rely on your attestation, and without it the agent refuses to read. - Targets. Provisioning writes to PostgreSQL and directories.
- Secret stores.
env://,file://andk8s://references are built in. Azure Key Vault, AWS Secrets Manager and HashiCorp Vault need a provider plugin in a derived agent image, or a secret staged into a file or environment variable. - Agent releases. The signed release machinery exists; until the first agent release is published, Downloads says so and the image is built from source.
- Private deployment. A dedicated or self-hosted control plane is a Docker Compose scaffold available on request: no control-plane Helm chart, no SAML, no high-availability topology, and air-gapped sites mirror images manually. Signed license files are issued after the production signing-key ceremony.
- Identity. Single sign-on uses OpenID Connect (OIDC). SAML-only identity providers need an OIDC bridge, and SCIM provisioning is not implemented.
- SDKs and CLI. The packages are publication-ready but not yet published to PyPI or npm.
Residual risks
The residual-risk register lists what remains after the implemented controls, each with an owner and a trigger for re-review. In summary:
| Id | Residual risk |
|---|---|
| RR-1 | Metadata such as schema, table and column names and aggregate counts would be exposed if the control plane were compromised. |
| RR-2 | One platform key signs session and agent tokens; tokens are not revoked on key compromise, and short lifetimes bound the exposure. |
| RR-3 | A stolen agent access token works for up to five minutes; a stolen agent private key works until the agent is revoked. |
| RR-4 | API rate limits apply per replica; shared limiting relies on the network edge. |
| RR-5 | Output written directly by native libraries bypasses structured log scrubbing. |
| RR-6 | The audit trail is tamper-evident, not tamper-proof. |
| RR-7 | Egress content detectors are heuristics; a compromised agent could use covert channels. |
| RR-8 | Separation of duties is enforced per account; one person holding two accounts defeats it. |
| RR-9 | Commands are not signed at the application layer; a compromised control plane could queue jobs against sources you registered. |
| RR-10 | Job sandboxes are directories on the agent host, not operating-system isolation. |
| RR-11 | Read-only proof is automatic only for PostgreSQL, SQLite and file systems. |
| RR-12 | A few documented cross-tenant system paths (credential resolution, scheduler, payment webhooks) must stay correct. |
| RR-13 | Stored contact-form data would need escaping if an internal viewer were ever built. |
| RR-14 | Offline self-hosted licensing can be reset by someone with host root and database-owner access. |
| RR-15 | Known-vulnerable dependencies may be accepted temporarily under time-boxed, owned exceptions. |
| RR-16 | Website and console builds are not single signed artifacts. |
| RR-17 | Host clock steps affect timestamp checks. |
| RR-18 | A local attacker with write access to the agent's datasets directory can tamper with published files; consistent rewrites of certified versions are detected. |
Legal documents
The Terms of Service, Privacy Policy, EULA and Data Processing Addendum published under Legal are drafts pending approval by legal counsel. They are not in force until approved, and the pages say so.
Benchmarks
DataNivra publishes a reproducible benchmark harness rather than headline numbers. The only results recorded so far come from a local developer-machine run; they are for engineering use and are not a performance claim. See Trust.
Analytics
Website analytics (Google Analytics 4) are off by default and consent-first: nothing from Google loads until the site operator has configured it and you choose Allow analytics. Global Privacy Control and Do Not Track count as "no". Only an allow-list of events with enumerated values is sent — never form contents, emails or identifiers. The console has no analytics. See the Privacy notice.