Sovereignty and compliance

DataNature Atlas in depth: RBAC, ABAC, automated PII masking and a governed semantic layer

DataNature Atlas is the platform's governance module: catalog, classification, lineage, certified metrics, access policies and audit. Atlas combines role-based access (who you are) with attribute-based policies (what the data is and what you are cleared for), masks PII automatically from classifications, and enforces the same rules in SQL, cubes, dashboards and AI agents, so one set of controls supports GDPR, PIPEDA, Law 25, CCPA, HIPAA and sector rules.

Published . Last updated

On this page

Governance usually fails at the seams: a policy defined in one tool and forgotten in another, a masked column copied unmasked into a feature table, an agent with a service account that reads everything. DataNature puts governance in one module that spans every engine: DataNature Atlas. Atlas knows what the data is, decides who may see it, and enforces that decision wherever data is read.

What does DataNature Atlas do?

CapabilityWhat Atlas provides
CatalogEvery dataset, column, owner, description and certification, searchable
ClassificationPII and sensitivity tags, detected automatically and confirmed by stewards
LineageSource-to-metric traceability across ingestion, dbt runs, cubes and models
Semantic layerCertified cubes and KPIs that every dashboard, app and agent shares
Access policiesRBAC grants, attribute-based policies, row filters and column masks
EnforcementDecisions applied in interactive SQL, serving compute, cubes, BI, exports and Voir agents
Quality and freshnessTest results and freshness signals per dataset
AuditWho or what accessed which data, masked or not, under which policy

Atlas answers both "what is this and where did it come from?" and "may this person or agent see it, and in what form?"

How does RBAC work?

Role-based access control assigns permissions to groups that mirror your organization, and users inherit them. DataNature uses RBAC at two levels:

  • Page access: View, Use or Manage for each part of the platform, such as notebooks, semantic models or security settings.
  • Data access: grants on catalogs, schemas, tables and columns to groups.

RBAC is simple to reason about, and it is the right base. Its weakness is scale: with many regions, products and sensitivity levels, roles multiply until nobody can explain why a person has access.

How does ABAC add to RBAC?

Attribute-based access control decides from attributes instead of fixed role lists:

  • User attributes: department, region, legal entity, clearance level, purpose.
  • Data attributes: classification tags from Atlas, such as pii, pii.national_id, confidential, regulated.payment, and the region or entity a row belongs to.
  • Context: the surface (SQL, dashboard, agent) and the purpose declared for a twin.

One attribute policy replaces dozens of role grants. Examples of policies Atlas enforces:

Policy in plain languageEffect
Columns tagged pii.national_id are nulled unless the user's clearance is kycNational IDs hidden everywhere except for the KYC team
Columns tagged pii.identifier are hashed for analytics groupsCustomer IDs stay joinable but unreadable
Rows are visible only where row.region equals the user's region attributeOne policy covers every regional manager
Data tagged regulated.payment may not be exportedCSV export blocked for payment tables
A twin declared for marketing cannot read columns tagged confidential.creditPurpose limitation for agents

Because policies follow tags, a new table that Atlas classifies as containing national IDs is protected the moment it is classified, without anyone writing a new grant.

How is PII found and masked automatically?

  1. Detection: as data lands in Bronze, Atlas scans columns and samples for personal data patterns (names, national IDs, phone numbers, emails, account and card numbers, dates of birth) and proposes classification tags.
  2. Classification: data stewards confirm or adjust tags; confirmed tags are inherited through lineage, so a Silver or Gold column derived from a PII column carries the tag forward.
  3. Masking: Atlas applies the mask type tied to each tag and user attribute: null, redact, show first or last four, hash, year only, or exempt.
  4. AI guardrails: PII detection also runs on agent answers, redacting sensitive values that should not appear in a response.
  5. Evidence: every masked and unmasked access is logged with the policy that allowed it.

Masking in Bronze and Silver, not only Gold, matters: analysts with SQL access can reach every layer.

Why does the semantic layer belong in governance?

Because a wrong number is a governance failure too. Atlas holds certified semantic models: cubes with measures, dimensions, joins and time grains over Gold tables, and certified KPIs with owners and descriptions. Atlas applies row predicates per role to every cube query, whether it comes from a dashboard, an application over REST or a Voir agent over MCP. The result: every consumer computes "revenue" the same way, and sees only the revenue they are allowed to.

How are AI agents governed?

Agents are treated as principals with their own scope, not as a back door:

  • Each Voir twin has a declared purpose and a data scope (selected catalogs, schemas, tables, document partitions).
  • Atlas evaluates every tool call: SQL allowlists, document partitions and attribute policies are checked before a system is reached.
  • Answers pass through PII guardrails.
  • Actions that change systems wait for human approval.
  • Atlas logs the user, the twin, the tool call, the data touched and the answer.

What evidence does this give auditors?

  • A catalog of every dataset with owner, classification and lineage.
  • Policies written in terms auditors understand (classification, attribute, effect), not thousands of grants.
  • An access log showing who or what read which data, masked or not, and under which policy.
  • Lineage from any report or agent answer back to source.
  • All of it stored inside your jurisdiction, because Atlas runs in your deployment.

How does Atlas help with privacy laws?

ObligationHow Atlas supports it
Know where personal data is (GDPR records of processing, Law 25, DPDP)Automated PII detection and classification across every layer
Limit access to what is necessary (GDPR, GLBA, HIPAA minimum necessary)RBAC plus attribute-based policies, row filters and masks
Purpose limitationPurpose declared per Voir twin; policies block out-of-purpose columns
Respond to access and deletion requests (GDPR, CCPA, PIPEDA)Catalog and lineage locate a person's data in every derived table
Demonstrate accountability to regulatorsAudit log, lineage and policy documentation, kept in-jurisdiction
Govern AI (EU AI Act, OSFI E-23, SR 11-7)Lineage from data to model to decision, logged agent actions, PII guardrails

Atlas provides controls; compliance itself is assessed for each organization's programme.

Frequently asked questions

Should we use RBAC or ABAC?

Both. Use RBAC for coarse access (who can use which tools and catalogs) and ABAC for fine-grained rules that depend on data sensitivity, region or purpose.

Do masks follow data into derived tables?

Classification tags are inherited through lineage, so derived columns carry the tag and the same masking policy applies.

Can policies be tested before they go live?

Yes. Disable a policy rather than deleting it while testing, test with a non-admin account because admins bypass many rules, and check the audit log after go-live. Changes take effect in about 30 to 60 seconds.

Sources

See it on your own data.

Book a demo

Demos and quotes are handled by Dlytica, the company behind DataNature.