SnowPro Specialty: Gen AI Question 217
Single answerLimits on which models can be usedA financial services company is deploying a Snowflake-based GenAI assistant for analysts. Due to internal governance, the platform team must ensure that analysts can invoke only an approved set of models and are prevented from using newer or unapproved models that may become available later. Which approach best satisfies this requirement in Snowflake while minimizing ongoing administrative overhead?
- A
Grant analysts broad privileges to use Cortex AI functions and rely on application code to reject prompts that specify unapproved models
- B
Create a network policy that restricts access to Snowflake endpoints for model invocations and assign it to analyst users
- C
Use a Cortex allowlist policy so only explicitly approved models can be used by the relevant users or roles
- D
Disable all external access integrations in the account so analysts cannot reach non-approved model providers
Show answer and explanation
Correct answer: C
Explanation
The best answer is to use a Cortex model allowlist policy so that only approved models can be invoked. This is the most direct way to enforce limits on which models can be used in Snowflake and is preferable to application-only checks, which are easier to bypass and harder to govern consistently. Network policies and external access integrations address different security concerns and are not model-governance controls for Cortex. In practice, Snowflake governance for GenAI should follow least-privilege principles: grant only necessary access and explicitly constrain model usage where account or organizational policy requires it. Candidates should recognize that the requirement is not merely about user access to AI features in general, but about restricting access to a defined subset of models in a maintainable way.
- A. Incorrect.
This is incorrect because broad access to Cortex AI functions does not itself restrict model choice. Relying on application code alone is weaker from a governance perspective because users may still access functions through other SQL paths, worksheets, notebooks, or applications if privileges permit. The requirement is to enforce approved-model usage at the Snowflake control plane level, not just in custom client logic.
- B. Incorrect.
This is incorrect because network policies control client access to Snowflake based on network location, not which Cortex models a user can invoke. They are useful for restricting login origins or endpoint access patterns, but they do not provide model-level governance for Snowflake Cortex model selection.
- C. Correct.
This is correct because an allowlist-style control for Cortex models is the governance mechanism intended to limit which models users can access. By explicitly approving models, the company prevents use of non-approved models, including newly introduced ones, unless they are added to the allowlist. This aligns with least-privilege principles and reduces operational overhead compared with trying to maintain blocking logic in each application.
- D. Incorrect.
This is incorrect because external access integrations govern outbound connectivity from Snowflake code such as UDFs or procedures to external services. They do not control which Snowflake Cortex models users may select. Disabling them could affect unrelated workloads and still would not be the correct control for limiting model usage inside Cortex.