SnowPro Specialty: Gen AI Question 17
Single answerSecurity, privacy, access, and control principlesA healthcare company is building an internal GenAI assistant in Snowflake to help support agents summarize case notes. The source notes contain sensitive patient information, and the company must ensure that support agents only see records for patients assigned to their region. The security team also wants to minimize exposure of raw sensitive text to downstream users while still allowing the application to generate summaries. Which approach BEST meets these requirements using Snowflake security and governance capabilities?
- A
Create a secure view over the case notes table that applies row access policies by region and masking policies to sensitive columns, and have the GenAI application query that governed view instead of the base table.
- B
Grant the application role direct SELECT access on the base case notes table, and rely on the LLM prompt instructions to avoid returning sensitive fields that the user should not see.
- C
Replicate the case notes into a separate table for each region, then let the application choose the correct regional table at runtime without using masking or row-level governance.
- D
Use a standard view on top of the base table and restrict access by giving support agents access only to a smaller warehouse so fewer rows are processed.
Show answer and explanation
Correct answer: A
Explanation
The best answer is to enforce security and privacy controls in Snowflake using policy-based governance on the data access path the GenAI application actually uses. For this scenario, row access policies address the requirement that agents only see patients assigned to their region, while masking policies help minimize exposure of sensitive patient data. Using a secure view provides an additional layer for governed consumption and is a common best practice when exposing protected data to applications. This follows core Snowflake security principles such as least privilege, centralized access control, and avoiding reliance on application-layer or prompt-layer behavior for enforcement. Relevant Snowflake guidance includes documentation on row access policies, dynamic data masking, secure views, and role-based access control (RBAC).
- A. Correct.
Correct. This is the strongest governance-oriented design. A secure view helps reduce unnecessary exposure of underlying logic and data patterns compared to directly exposing the base table. Applying a row access policy enforces region-based row-level filtering so users only access authorized patient records. Applying masking policies to sensitive columns reduces exposure of protected information while still allowing governed access for summarization workflows. Having the GenAI application read from the governed view aligns with least-privilege access and centralized policy enforcement.
- B. Incorrect.
Incorrect. Prompt instructions are not a security boundary. If the application role has direct access to the base table, sensitive data is already exposed to the application query path. Snowflake governance should be enforced with access controls and policies, not delegated to model behavior. This option reflects a common misconception that LLM prompting can replace data access controls.
- C. Incorrect.
Incorrect. Creating separate physical tables by region can increase operational complexity, duplication, and risk of inconsistent governance. It does not address masking of sensitive fields and is generally less maintainable than centralized policy-based controls such as row access policies and masking policies. It may work functionally, but it is not the best approach for scalable, governed access control.
- D. Incorrect.
Incorrect. Warehouse size or access to a smaller warehouse does not enforce data security. Compute resources control performance and cost, not row-level authorization or column-level privacy. A standard view alone also does not provide the same governance intent as explicit row access and masking policies.