Snowflake Security, RBAC, and User Administration · Topic 8 of 16
Account-Level Permissions and Access Control for AI and ML Models
Manage Snowflake account-level privileges such as MANAGE GRANTS and CREATE INTEGRATION and secure access to ML models and Cortex AI functions.
- 4 min
- MANAGE GRANTS, CREATE INTEGRATION, ATTACH POLICY, MONITOR USAGE and MANAGE WAREHOUSES are account level privileges.
- Give them to dedicated custom roles, not to broad roles.
- Models need CREATE MODEL (schema), USAGE or READ, and ownership for artifacts.
What it is
Some privileges belong to the account itself, not to a database or schema. These global privileges let a role grant access, create integrations, attach policies or monitor usage. They are powerful, so an administrator grants them to custom roles one at a time and checks who holds them.
AI and ML objects follow the same model. Models in the Snowflake Model Registry are schema objects with their own privileges, and Snowflake Cortex AI functions are controlled by an account privilege and database roles.
How it works
Account-level privileges
| Privilege | Allows |
|---|---|
| MANAGE GRANTS | Grant or revoke privileges on objects the role does not own; set future grants |
| CREATE INTEGRATION | Create any integration type; must be granted by ACCOUNTADMIN |
| CREATE ACCOUNT | Create a managed (reader) account; must be granted by ACCOUNTADMIN |
| ATTACH POLICY | Activate a network policy for the account |
| APPLY MASKING POLICY | Set masking policies on columns and tags |
| MONITOR USAGE | Monitor account level usage and history |
| MONITOR EXECUTION | Monitor pipes and tasks |
| MANAGE WAREHOUSES | MODIFY, MONITOR and OPERATE on all warehouses |
| EXECUTE TASK | Run tasks owned by the role |
| IMPORT SHARE | View shares; creating a database from a share also needs CREATE DATABASE |
| USE AI FUNCTIONS | Use Cortex AI functions (with a database role) |
USE ROLE ACCOUNTADMIN;
GRANT MANAGE GRANTS ON ACCOUNT TO ROLE grants_admin;
GRANT CREATE INTEGRATION ON ACCOUNT TO ROLE integration_admin;
SHOW GRANTS ON ACCOUNT;
Practices:
- Grant each privilege to a dedicated custom role and roll that role up to the right system role. Do not mix account management privileges with object privileges.
- Review holders with
SHOW GRANTS ON ACCOUNTandACCOUNT_USAGE.GRANTS_TO_ROLES. ALTER ACCOUNTparameters (for example, network policy or session policy at account level) need ACCOUNTADMIN or a role with the matching privilege.- Treat MANAGE GRANTS like ownership of everything: it lets a role grant itself access to objects.
A role with MANAGE GRANTS can give itself any object privilege. Keep the number of holders small and audit them.
Access control for ML models
Models in the Snowflake Model Registry are schema objects. The docs list these rules:
- Creating a model needs the
CREATE MODELprivilege on the schema, or schema ownership. USAGEruns the model for warehouse inference without seeing internals.READruns the model for container inference and allows viewing metadata such as comments, tags and metrics.- Viewing or downloading model artifacts needs ownership.
GRANT CREATE MODEL ON SCHEMA ml.registry TO ROLE ml_engineer;
GRANT USAGE ON ALL MODELS IN SCHEMA ml.registry TO ROLE analyst;
GRANT USAGE ON FUTURE MODELS IN SCHEMA ml.registry TO ROLE analyst;
The usual pattern is: data scientists own models and hold CREATE MODEL, consumers hold USAGE, and future grants keep newly created models reachable.
Access control for Cortex AI functions
| Item | Behavior |
|---|---|
USE AI FUNCTIONS account privilege | Needed with a database role; granted to PUBLIC by default; only ACCOUNTADMIN manages it |
SNOWFLAKE.CORTEX_USER database role | Enables calls to AI functions; granted to PUBLIC by default |
SNOWFLAKE.AI_FUNCTIONS_USER | Scalar AI functions only; not granted to PUBLIC by default |
SNOWFLAKE.CORTEX_EMBED_USER | Embedding functions; not granted to PUBLIC by default |
REVOKE USE AI FUNCTIONS ON ACCOUNT FROM ROLE PUBLIC;
GRANT USE AI FUNCTIONS ON ACCOUNT TO ROLE analyst;
GRANT DATABASE ROLE SNOWFLAKE.CORTEX_USER TO ROLE analyst;
Access to specific models is covered with the AI access topics later in this chapter.
To restrict AI to one team, revoke from PUBLIC first, then grant to a custom role. Granting without revoking changes nothing, because PUBLIC still has it.
Common mistakes
Granting MANAGE GRANTS widely. It is an escalation path.
Granting CREATE INTEGRATION without ACCOUNTADMIN. Only ACCOUNTADMIN can grant it.
Giving USAGE where READ is meant. USAGE is warehouse inference; READ is container inference and metadata.
Granting AI access without revoking PUBLIC. The restriction has no effect.
What to remember for the exam
- MANAGE GRANTS, CREATE INTEGRATION, ATTACH POLICY, MONITOR USAGE and MANAGE WAREHOUSES are account level privileges.
- Give them to dedicated custom roles, not to broad roles.
- Models need CREATE MODEL (schema), USAGE or READ, and ownership for artifacts.
- AI functions need USE AI FUNCTIONS and CORTEX_USER or AI_FUNCTIONS_USER; PUBLIC holds USE AI FUNCTIONS and CORTEX_USER by default.
- Revoke from PUBLIC before granting to a team.
Check yourself
Which privilege allows creating integrations?
CREATE INTEGRATION, granted by ACCOUNTADMIN.Which model privilege runs inference without exposing artifacts?
USAGE, for warehouse inference.How do you limit Cortex AI functions to one role?
Revoke
USE AI FUNCTIONS(and the CORTEX_USER role) from PUBLIC and grant them to that role.
Topic 8 of 16 in Snowflake Security, RBAC, and User Administration
Test yourself on Snowflake Security, RBAC, and User Administration
The full Snowflake ADA-C02 question bank has 814 questions, with timed practice exams and an explanation for every answer.
This study guide is independent and is not affiliated with, endorsed by, or sponsored by Snowflake. Vendor names, certification names and exam codes are trademarks of their respective owners and are used for identification only.