GUIDE, read 13 Nov 2026
Snowflake enforces a sensitive field at query time, and its own record stops where the BI tool begins
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.
Snowflake's own controls are strongest between the column and the query
Snowflake says its classification feature discovers sensitive data automatically and lets you attach tags and masking policies to it. Per Snowflake's docs, a masking policy set on a tag protects every tagged column of a matching data type, and a policy attached to a tag is applied when classification applies that tag. Row access policies add row-level filtering on Enterprise Edition. Snowflake Horizon holds the column stage at documented and the agent stage at documented in the exposure map.
Snowflake says that when an AI agent queries data, the role's masking and row access policies are enforced in the query engine, the same as for a person. The managed MCP server adds its own grants: access to the server does not grant access to a tool, and each tool needs its own grant.
The stage table shows where Snowflake's record ends
Read left to right: each row is one stage of the path a sensitive column takes, with what Snowflake's features do and where they stop.
| Stage | What Snowflake's own features do | Where it stops |
|---|---|---|
| Column | Sensitive data classification writes SEMANTIC_CATEGORY and PRIVACY_CATEGORY tags; tag-based masking and row access policies act on them | Classification skips tables with more than 10,000 columns, and AI mode skips VARIANT, OBJECT and ARRAY columns |
| Transformation | Tag propagation copies a tag from a source table to views and tables built on it (Enterprise Edition) | Tags set through CTAS, INSERT, MERGE or COPY INTO are not updated afterward; propagation fails past 10,000 downstream objects |
| BI field | Horizon connectors read workbook and field metadata from Tableau and Power BI | None found on 13 Nov 2026: sensitivity tags carried onto BI fields; the Horizon Context page marks these connectors Preview |
| Report | Connectors collect view and workbook usage and popularity | Popularity is a count, not a list of people |
| Viewer | ACCESS_HISTORY records which user touched which columns, for 365 days | Latency up to 180 minutes; none found on 13 Nov 2026: named viewers per BI report |
| Agent | Masking and row policies run in the engine; each MCP tool needs its own grant | SYSTEM_EXECUTE_SQL on the same server lets the client bypass the agent's semantic views |
Classification writes system tags, and Snowflake says system tags do not propagate
This is the detail most teams find late. Snowflake's classification page calls SNOWFLAKE.CORE.SEMANTIC_CATEGORY and SNOWFLAKE.CORE.PRIVACY_CATEGORY system-defined tags. Its tag propagation page lists "System tags are not propagated" under limitations. Read together, a freshly classified column does not pass its classification tag to a view on its own. The documented route is a tag map: the classification page shows a user-defined tag (such as a pii tag with the value Highly confidential) applied whenever a given system tag is applied, and user-defined tags can propagate.
The same limitations list warns that a tag tied to a masking policy can sit on both a source table and a target view, so the policy may execute twice. Snowflake documents a tag removed at the source being removed from the target, which keeps a withdrawn label from lingering. Test this order in a non-production account before you rely on it, since the two pages are written separately.
ACCESS_HISTORY answers who read a column, and a BI service account hides the viewer
ACCESS_HISTORY lists the user, the objects and columns accessed, and a policies_referenced field showing which masking or row access policy applied to the query. Snowflake documents latency of up to 180 minutes, so an incident review at hour one will miss recent reads. On 20 Oct 2026 Snowflake made two Trust Center reports generally available: one lists the users and roles with privileges on tables holding sensitive data, the other lists users who accessed those tables in a chosen lookback period. Both need Enterprise Edition.
Where a BI tool connects with one shared service user, the user in these records is that account. This is a reasoned reading, not a Snowflake statement: the person who opened the dashboard sits in the BI tool's own log, which is why the viewer cell reads partial.
Which ranked tools extend it past the warehouse
Four kinds of tool carry the field beyond Snowflake. Column-level lineage into BI fields is documented by Alation for Tableau and Power BI (documented at the BI field stage), by Select Star for Looker and Sigma (documented), and by Secoda for Looker and Tableau only (partial). Atlan documents column-level lineage from Tableau datasource to worksheet to dashboard and syncs Snowflake tags into its catalog; its tag propagation is off by default and chosen per tag.
Euno is a context platform that reads metadata and query logs only. Per Euno's docs, it takes classifications from Cyera and BigID, maps them onto Snowflake resources, and traces them forward through column-level lineage into BI fields and reports. It ranks 1 of 16 on this site; its BI field cell reads documented and its viewer cell partial.
For enforcement outside Snowflake's own policies, Immuta says it ingests Snowflake table and column tags and authors masks by tag; its Tableau setup needs a live connection, and viewers see data under their own policies when the published data source prompts for credentials. Snowflake's reach at the viewer is partial.
Check four things before an agent connects to Snowflake data
First, query SNOWFLAKE.CORE.PRIVACY_CATEGORY on the columns you care about and confirm a user-defined tag mapped to it is the one your masking policy uses. Second, expose a Cortex Agent as the client-facing tool and keep SYSTEM_EXECUTE_SQL on a separate server with a least-privileged role, as Snowflake recommends. Third, ask your BI owner whether the Tableau or Power BI connection uses per-user credentials, because that decides what ACCESS_HISTORY can say. Fourth, note the 180-minute latency in any audit runbook.
Sources
- Snowflake: sensitive data classification and system tags, read 13 Nov 2026
- Snowflake: automatic tag propagation and limitations, read 13 Nov 2026
- Snowflake: ACCESS_HISTORY view, read 13 Nov 2026
- Snowflake: managed MCP server and grants, read 13 Nov 2026
- Snowflake: tag-based masking policies, read 13 Nov 2026
- Snowflake: Horizon Context, read 13 Nov 2026
- Snowflake release note, 20 Oct 2026: Trust Center sensitive data reports, read 13 Nov 2026
- Atlan: what it crawls from Tableau, read 13 Nov 2026
- Euno docs: Cyera integration, read 13 Nov 2026
- Immuta docs: Tableau, read 13 Nov 2026
Last reviewed 13 Nov 2026.