ADA-C01 Question 65
Single answerEnable security and access control for AI/ML modelsA financial services company is rolling out a Snowflake Cortex AI feature that lets analysts summarize internal support cases stored in a table containing sensitive customer information. The security team requires that only members of the ANALYST_AI role can invoke the model from a worksheet, while the underlying case data must remain inaccessible to other users. In addition, administrators want to avoid granting broad account-level privileges that would let users invoke every available model unnecessarily. Which action best meets these requirements?
- A
Grant the ANALYST_AI role USAGE on the specific Cortex model object and SELECT on the source table, then rely on warehouse-level permissions to limit who can run inference.
- B
Grant the ANALYST_AI role the database role SNOWFLAKE.CORTEX_USER and only the minimum object privileges required to read the approved source data.
- C
Grant the ANALYST_AI role ACCOUNTADMIN temporarily, let users create secure views over the source data, and revoke ACCOUNTADMIN after the views are created.
- D
Grant the ANALYST_AI role IMPORTED PRIVILEGES on the SNOWFLAKE database and assume this limits access to only approved AI functions automatically.
Show answer and explanation
Correct answer: B
Explanation
The best answer is to grant the custom role the SNOWFLAKE.CORTEX_USER database role and separately grant only the minimum data-access privileges needed for the approved workflow. This follows Snowflake security best practices: use role-based access control, grant least privilege, and separate access to AI features from access to underlying business data. In practice, administrators should create a dedicated custom role for AI-enabled analysts, grant that role the appropriate Cortex database role from the SNOWFLAKE database, and then grant SELECT only on the specific tables or secure views that contain approved data. This avoids overprivileged account roles such as ACCOUNTADMIN and avoids treating Cortex access as if it were controlled by generic warehouse permissions or object-level model grants. Candidates should know that Snowflake documents feature access for Cortex through Snowflake-provided database roles and recommends using custom roles to encapsulate business-specific access patterns.
- A. Incorrect.
Incorrect. Snowflake Cortex access is not managed by granting USAGE on an individual model object in the same way as regular schema objects. This option also overemphasizes warehouse permissions, which control compute usage but do not provide the required AI feature authorization model. While SELECT on the source table may be necessary for direct access, the key security requirement is granting the appropriate Cortex access role rather than model-object USAGE.
- B. Correct.
Correct. Access to Snowflake Cortex LLM functions is typically enabled through the SNOWFLAKE.CORTEX_USER database role, which can be granted to a custom role such as ANALYST_AI. This aligns with least-privilege administration by avoiding broad account-level roles. Users still need separate privileges, such as SELECT on approved tables or views, to access the underlying data used in prompts or summarization workflows. This combination satisfies both AI feature access control and data access control requirements.
- C. Incorrect.
Incorrect. Granting ACCOUNTADMIN, even temporarily, violates least-privilege principles and introduces unnecessary risk. Creating secure views can help restrict data exposure, but ACCOUNTADMIN is not the appropriate mechanism for enabling model access for analysts. The exam expects candidates to recognize that broad administrative roles should not be used when scoped feature-access roles are available.
- D. Incorrect.
Incorrect. IMPORTED PRIVILEGES on the SNOWFLAKE database is not the right control for scoped Cortex user access in this scenario. The intended pattern is to grant the relevant SNOWFLAKE database role, such as SNOWFLAKE.CORTEX_USER, to a custom role. Assuming IMPORTED PRIVILEGES will automatically constrain users to approved AI functions reflects a misunderstanding of how privileges to shared system databases differ from feature-specific database roles.