Question 370
At fictional Brookstream Commerce Alliance, a Snowflake data and AI engineering team is conducting a restricted enterprise language-model deployment using approved synthetic testing data instead of confidential customer records. The separately composed technical incident is: A customer analysis team sends full identity and payment fields to an LLM even though the classification goal needs only transaction categories and anonymized descriptions. Assume only these documented facts and use safe, authorized production-change procedures; exact current Snowflake feature syntax and region or account availability require current review. Which repeatable assurance or control is most relevant to preventing this situation?
Explanation
Writer-proposed answer: B (UNVERIFIED)
Preliminary individual option explanations — NOT expert validated
Correct Answer
Embed data classification and masking in pipeline policies, test payload samples and reauthorize new attributes before use. Data minimization should occur before inference because downstream redaction cannot undo an already disclosed sensitive prompt.
Incorrect Answers
This response concerns Agent tool execution requiring approvals, which is different from the stated failure. A customer analysis team sends full identity and payment fields to an LLM even though the classification goal needs only transaction categories and anonymized descriptions.
Overprivileged AI service account roles is not the cause of the reported failure: A customer analysis team sends full identity and payment fields to an LLM even though the classification goal needs only transaction categories and anonymized descriptions.
Cortex model access allowlist governance alone does not establish the needed evidence: Sampled traces show required business classifications still work while unnecessary personal fields never enter model requests or retained logs.