SnowPro Specialty: Gen AI Question 240
Single answerCortex GuardA financial services company is building a customer-support chatbot in Snowflake using Cortex AI functions. During testing, the team discovers that some users can craft prompts that try to override system instructions and ask the model to reveal internal policies or produce disallowed content. The security team wants a control that can inspect prompts and model responses for unsafe or policy-violating content before the application returns output to users. Which approach best addresses this requirement?
- A
Use Cortex Guard to evaluate input prompts and generated responses against safety policies, and block or handle flagged interactions in the application flow
- B
Increase the model temperature so that unsafe prompts are less likely to produce deterministic policy-violating outputs
- C
Grant fewer Snowflake object privileges to end users so they cannot access the model's unsafe responses
- D
Use a larger warehouse so content safety checks can be enforced through query execution controls
Show answer and explanation
Correct answer: A
Explanation
The key requirement is to detect and handle unsafe or policy-violating interactions in a GenAI application, particularly prompt injection or jailbreak-style attempts and problematic generated outputs. Cortex Guard is the control designed for this type of AI safety use case. In practice, teams place such checks around both inbound prompts and outbound responses so the application can block, redact, escalate, or otherwise handle flagged content according to policy. By contrast, model hyperparameters such as temperature affect generation characteristics, not compliance or safety guarantees. Likewise, Snowflake RBAC and warehouse sizing are important for platform operations and governance, but they do not inspect LLM content. Best practice for enterprise GenAI applications is to combine least-privilege access controls with dedicated content-safety guardrails such as Cortex Guard.
- A. Correct.
Correct. Cortex Guard is intended to help detect unsafe or policy-violating content in GenAI workflows. In this scenario, the requirement is to inspect both user inputs and model outputs for issues such as jailbreak attempts, prompt injection, or disallowed content before returning a response. Applying Cortex Guard in the request/response flow is the most appropriate control because it directly addresses LLM safety and moderation requirements.
- B. Incorrect.
Incorrect. Temperature affects response randomness and creativity, not safety enforcement. Lowering or raising temperature does not provide a reliable mechanism for detecting prompt injection, jailbreak attempts, or policy violations. A candidate might choose this because model parameter tuning can influence output style, but it is not a substitute for a safety guardrail.
- C. Incorrect.
Incorrect. Snowflake access control governs access to data and objects, not semantic inspection of prompts or generated text. Even with tightly scoped privileges, a model can still receive unsafe instructions or generate harmful content unless a content-safety control evaluates the interaction. This option confuses platform authorization with LLM safety filtering.
- D. Incorrect.
Incorrect. Warehouse size affects compute resources for query execution, but it does not provide safety moderation or policy enforcement for LLM inputs and outputs. Someone might pick this if they associate all Snowflake processing controls with warehouses, but Cortex Guard addresses a different problem: content risk and policy checks in GenAI applications.