SnowPro Advanced: Security Engineer Question 423
Single answerInterpret traces to debug and audit the flow of sensitive data through Gen AI applicationsA security engineer is investigating whether personally identifiable information (PII) from customer support chats is being exposed to an LLM inside a Snowflake-hosted Gen AI application. The team enabled observability for the app and captured traces for each request. During review, the engineer needs to reconstruct the request path, identify which component introduced the sensitive value, and preserve evidence for audit without broadly exposing prompt contents to other users. Which action is the MOST appropriate?
- A
Query the recorded traces to follow parent-child span relationships across the application workflow, inspect span attributes/events for the sensitive field, and restrict access to the trace data using Snowflake RBAC so only authorized investigators can review it.
- B
Rely on QUERY_HISTORY alone to determine where the PII entered the LLM prompt, because SQL text fully captures the end-to-end Gen AI request flow across application components.
- C
Disable tracing after the incident and use warehouse load history to infer which service step added the sensitive value, since observability traces are intended only for performance tuning.
- D
Grant broad read access to the trace repository for all developers so they can search prompt payloads directly, because faster access is more important than limiting exposure during an audit.
Show answer and explanation
Correct answer: A
Explanation
The best answer is to use observability traces as the primary source for reconstructing the end-to-end flow of a Gen AI request, then correlate spans to determine where sensitive data was introduced. In practice, traces are designed to show execution across components, making them more suitable than SQL-only artifacts for debugging and auditing Gen AI workflows. From a security perspective, trace records can themselves become sensitive if they include prompt fragments, model inputs, or other metadata. Therefore, best practice is to protect trace access with Snowflake RBAC and least-privilege controls, preserving evidence for audit while minimizing unnecessary exposure. This aligns with Snowflake security principles around role-based access control, governance, and controlled access to sensitive operational data.
- A. Correct.
Correct. To debug and audit sensitive-data flow in a Gen AI application, the engineer should use traces to reconstruct the execution path through spans and their relationships. Parent-child span relationships help identify which component or step introduced the sensitive value. Reviewing span metadata, attributes, and events is the right way to pinpoint where data appeared. Because trace data may itself contain sensitive information, access should be limited through Snowflake role-based access control and least-privilege practices so only authorized personnel can investigate.
- B. Incorrect.
Incorrect. QUERY_HISTORY can help with SQL-level activity, but it does not provide full distributed tracing across all Gen AI application components. It is not sufficient for reconstructing an end-to-end request path or identifying which non-SQL application step introduced a sensitive value into a prompt. This option reflects the common misconception that SQL history alone replaces observability traces.
- C. Incorrect.
Incorrect. Warehouse load history is unrelated to determining the logical path of sensitive data through a Gen AI workflow. Tracing is not only for performance analysis; it is also valuable for debugging, auditing, and understanding request execution paths. Disabling tracing during or immediately after an incident would also reduce the ability to preserve and analyze evidence.
- D. Incorrect.
Incorrect. Broad access to trace data violates least-privilege security principles, especially when prompts or trace attributes may contain sensitive content. Even during an incident, access should be tightly controlled. Making traces widely available increases the risk of further exposure rather than improving secure investigation.