# dbt records a sensitive column by hand and does not carry the mark from a source column into the models

GUIDE, read 13 Nov 2026

**The short answer.** dbt holds sensitivity as hand-written tags and meta on columns, and its docs say a column's tags and meta do not inherit from the parent resource. Grants run at build time and masking is left to warehouse policies. Of the 16 ranked tools, ten name dbt among the transformation tools they read.

**Where Euno is behind.** Euno is not a policy engine or a DSPM: it reads metadata only, so it does not scan data contents or mask at the warehouse, which leaves it behind Atlan on acting on findings (5 to 7), and it publishes no dollar prices.

## dbt marks sensitivity with tags and meta that a person writes

dbt's reference docs show the pattern directly: a source column is designated as containing PII by setting a meta key such as contains_pii to true in YAML,; the tags config can mark whole models the same way. Automatic detection of sensitive columns in dbt's docs: none found on 13 Nov 2026. A person decides, types the key and keeps it current.

That makes the mark as good as the process around it. A column added next sprint carries no mark until someone adds one, and nothing in dbt's docs raises a flag when the warehouse's own classification disagrees with the YAML.

## A column's tags and meta stop at the column

dbt's columns reference says columns are not resources, so their tags and meta are not true configurations. In its words, they do not inherit the tags or meta values of their parent resources. The same reasoning extends across models: a source column marked contains_pii true does not hand that mark to the model column that selects it. Each model that exposes the value needs its own entry, or a tool outside dbt has to read the lineage and copy the mark.

Two related limits are easy to miss. dbt's meta docs say changes to meta, including at the column level, do not trigger state:modified, because dbt treats meta and tags as metadata only. A pull request that adds or removes a PII flag will not, by that rule alone, select the model in a state-based CI run. And the persist_docs config, which writes descriptions into the warehouse as comments, is documented as not implemented for sources, and it writes descriptions, not meta keys. A PII flag kept only in meta never reaches the warehouse through that route.

## Contracts and access control structure, not readers

A model contract is checked at build time: a model that does not match its declared columns and types fails to build. Contracts, access and versions apply to models only, not to sources, seeds or snapshots, so the layer where raw sensitive columns land is outside them. Contracts can declare constraints such as not_null, but the docs describe them as column shape, not as sensitivity.

The access modifiers (private, protected, public) govern who may ref() a model: the same group, the same project, or anyone. dbt's docs state that model access is not user access. They also say that when a macro is invoked with dbt run-operation, dbt does not enforce model access or group controls. An approver who reads private as hidden from analysts has read a feature that governs project references, not warehouse queries.

## Grants and masking happen in the warehouse, and dbt only issues them

The grants config applies privileges at build time: when a model finishes building, dbt ensures the grants on the table or view match the configuration exactly, and the grants are included in the manifest. That gives a reviewable record of who can select from each model. It works at table level.

For row- and column-level access and masking policies, dbt's grants page points to hooks and says to write the statements manually. The policy itself lives in the warehouse. For Databricks, dbt Labs documents a databricks_tags config (dbt-databricks 1.11 or higher) that applies Unity Catalog tags to materialized views and streaming tables, which is one route for a model-level mark to reach the platform's tag-based controls. Immuta's docs describe a related timing issue: its policies attach by tag to tables created by dbt, but it does not receive DDL events from the data platform, which leaves a window it calls policy downtime that its documented steps are meant to shorten.

## Docs and exposures describe consumers without describing sensitivity

dbt exposures declare downstream uses such as dashboards, with an owner and a maturity of high, medium or low. Exposures carry no sensitivity field. An exposure tells a reviewer that a dashboard depends on a model. Whether the model holds a regulated value is back to the hand-written tag.

## Ten ranked tools read dbt and some carry the mark further

The tools that list dbt among the transformation tools they read are Euno, Collibra, Atlan, Alation, Ataccama ONE, Databricks Unity Catalog, Snowflake Horizon, DataHub, Select Star and Secoda. The rest of the 16 name none: Informatica, Immuta, Privacera, Securiti, OneTrust and Microsoft Purview list no dbt connector in their evidence files. On the exposure map, the transformation stage reads partial for Euno, partial for Atlan, partial for Collibra, partial for Alation, partial for Ataccama ONE, partial for Databricks, partial for Snowflake, partial for DataHub, partial for Select Star and partial for Secoda.

What they add differs. Atlan says it crawls and syncs tags from sources including dbt. Select Star says dbt tags are synced and can be matched to its own tags, with PII tags as an example. Euno's AlphaSense write-up describes a rule that flags anything depending on a PII column upstream as not AI-ready, which reads upstream dependency instead of a hand-written tag on the downstream model. See [the transformation column](/exposure-map/transformation/) and the [classification guide](/guides/classification-from-warehouse-to-agent/) for the full set.

## Checks before you rely on a dbt tag for a regulated field

Count the models that select a tagged source column and see how many carry their own tag. Run a CI check that fails a pull request when a PII source column is selected by an untagged model column. Confirm the warehouse policy, not the YAML, is what hides the value. Ask each tool you use which mark it trusts when dbt and the warehouse classification disagree.

## Sources

- dbt docs: meta: https://docs.getdbt.com/reference/resource-configs/meta (read 13 Nov 2026)
- dbt docs: columns: https://docs.getdbt.com/reference/resource-properties/columns (read 13 Nov 2026)
- dbt docs: persist_docs: https://docs.getdbt.com/reference/resource-configs/persist_docs (read 13 Nov 2026)
- dbt docs: model contracts: https://docs.getdbt.com/docs/mesh/govern/model-contracts (read 13 Nov 2026)
- dbt docs: model access: https://docs.getdbt.com/docs/mesh/govern/model-access (read 13 Nov 2026)
- dbt docs: grants: https://docs.getdbt.com/reference/resource-configs/grants (read 13 Nov 2026)
- dbt docs: Databricks configs: https://docs.getdbt.com/reference/resource-configs/databricks-configs (read 13 Nov 2026)
- dbt docs: exposures: https://docs.getdbt.com/docs/build/exposures (read 13 Nov 2026)
- Immuta: dbt transform workflow: https://documentation.immuta.com/saas/govern/authoring-policies-in-secure/dbt-transform-workflow.md (read 13 Nov 2026)
- Atlan: tags: https://docs.atlan.com/llms/governance/tags/what-are-tags/llms.txt (read 13 Nov 2026)
- Select Star: docs: https://web.archive.org/web/20260715110406/https://docs.selectstar.com/llms-full.txt (read 13 Nov 2026)
- Euno: AlphaSense and dbt write-up: https://euno.ai/blog/alphasense-x-euno-dbt-summit-26 (read 13 Nov 2026)

Last reviewed 13 Nov 2026. Canonical: https://aidatagovernancetools.com/guides/sensitive-fields-in-dbt/
