SnowPro Specialty: Gen AI Question 28
Single answerControl model accessA financial services company is rolling out Snowflake Cortex capabilities to multiple business units. The security team requires that only the Data Science role can use a specific set of approved LLMs for inference, while Analyst roles must not be able to invoke those models even if they can run SQL in the same database and schema. The team wants a solution that follows Snowflake's access control model for AI features rather than relying only on application-side checks. Which action should the Snowflake administrator take?
- A
Grant the necessary Cortex model access privileges only to the Data Science role, and do not grant those model access privileges to Analyst roles.
- B
Grant USAGE on the database and schema to the Analyst roles, because model invocation is controlled entirely by object-level SQL privileges.
- C
Create a masking policy on prompt columns so Analyst roles cannot access the models unless the masking policy evaluates to TRUE.
- D
Allow Analyst roles to access the approved models, but restrict the maximum token count in their SQL statements to prevent actual inference.
Show answer and explanation
Correct answer: A
Explanation
The best answer is to use Snowflake role-based access control to grant access only to the approved models for the Data Science role. In Snowflake, access to AI capabilities such as Cortex models should be governed through platform permissions, not just application logic or generic SQL object grants. This is the most secure and auditable approach because it enforces least privilege at the platform level. Standard privileges on databases, schemas, or tables may still be required for related data access, but they do not substitute for explicit authorization to use specific models. Masking and token-related controls can complement a security design, but they do not satisfy the core requirement of restricting which roles can invoke particular models. This aligns with Snowflake best practices for centrally managing access to AI features and approved models using RBAC.
- A. Correct.
Correct. Controlling access to specific Cortex models should be done through Snowflake's role-based access control for model access, not only through database or schema privileges. If the requirement is that only the Data Science role can invoke certain approved models, the administrator should grant the relevant model access privileges to that role and withhold them from Analyst roles. This aligns with least-privilege access and centralizes enforcement in Snowflake rather than in client code.
- B. Incorrect.
Incorrect. Database and schema privileges such as USAGE are necessary for accessing SQL objects, but they do not by themselves authorize a role to invoke specific AI models. This option reflects a common misconception that standard SQL object permissions alone govern Cortex model access. In practice, model access is controlled separately and should be granted explicitly to the intended roles.
- C. Incorrect.
Incorrect. Masking policies are designed to protect sensitive data values in query results, not to govern whether a role can invoke a particular Cortex model. While masking may help reduce exposure of confidential prompt content, it does not replace model-level access controls. Someone might choose this because it sounds security-related, but it addresses data visibility rather than authorization to use a model.
- D. Incorrect.
Incorrect. Token limits are not the primary security mechanism for controlling which roles may invoke models. Even if a workload control or application convention limited token usage, Analyst roles would still have model access if permissions were granted. This option confuses consumption governance with authorization. The requirement is to prevent model invocation altogether for certain roles, which should be handled with RBAC.