SnowPro Specialty: Gen AI Question 233
Single answerIs data leaving/going to LLMs?A healthcare company is evaluating Snowflake Cortex AI functions to summarize support cases that may contain sensitive patient information. The security team asks whether using these LLM-powered functions means their data is sent outside Snowflake-managed infrastructure to third-party model providers. Which statement best answers this concern?
- A
All prompts and responses are always retained by the model provider for training, so Cortex AI functions should not be used with sensitive data.
- B
When using Snowflake Cortex AI functions, data can be processed by models hosted by Snowflake or by external model providers depending on the selected model, so the team should review the model and data handling details before use.
- C
Cortex AI functions guarantee that no data ever leaves the customer’s cloud region, regardless of model choice or feature configuration.
- D
If the company masks columns in a view before calling a Cortex AI function, Snowflake guarantees that no prompt data is sent to any LLM.
Show answer and explanation
Correct answer: B
Explanation
This question tests whether the candidate can distinguish between Snowflake data governance controls and the actual inference path used by LLM-powered services. The key principle is that teams must understand whether a given Snowflake Cortex AI feature uses a Snowflake-hosted model or an external model provider, because that affects whether data is processed only within Snowflake-managed infrastructure or sent to a provider for inference. For sensitive workloads, best practice is to review Snowflake documentation for the specific Cortex AI function or model, validate provider and regional behavior, minimize sensitive prompt content, and apply governance controls such as masking and role-based access before invoking AI functions. This aligns with Snowflake guidance to review feature-specific data handling details rather than assuming a universal answer for all LLM use cases.
- A. Incorrect.
Incorrect. This overstates the risk and is not an accurate description of Snowflake Cortex AI behavior. Snowflake provides controls and documentation about model usage and data handling, and it is not correct to say that prompts and responses are always retained by providers for training. A candidate might choose this if they assume all third-party LLM usage implies training on customer data, which is a common misconception.
- B. Correct.
Correct. This reflects the practical governance decision teams must make. In Snowflake Cortex AI, data handling depends on the model and service path involved. Some capabilities may use Snowflake-hosted models, while others may involve external model providers. Organizations should verify the specific model, region availability, and associated data handling guidance before sending sensitive content.
- C. Incorrect.
Incorrect. This is too absolute and ignores that data residency and processing paths can vary by service and model. A candidate might select this if they confuse Snowflake data platform controls with a blanket guarantee about every LLM inference path. For regulated workloads, teams must validate the exact behavior for the chosen Cortex feature and model.
- D. Incorrect.
Incorrect. Masking or transforming data before inference can reduce exposure, but it does not itself guarantee that no prompt data is sent to an LLM. If a Cortex AI function is invoked, the resulting prompt content used for inference is still processed by the selected model path. This option reflects the misconception that upstream masking changes where inference happens.