SnowPro Specialty: Gen AI Question 22
Single answerGuardrailsA financial services company is building a customer-support assistant in Snowflake using Cortex AI. The assistant answers questions about account features and troubleshooting steps based on approved internal knowledge. During testing, the team finds two recurring issues: users can sometimes manipulate prompts to make the model ignore the assistant's instructions, and some responses include language that violates the company's safety policy. The team wants a solution that reduces these risks without requiring them to manually review every prompt and response. Which approach is the MOST appropriate?
- A
Implement Cortex Guard to evaluate inputs and outputs for policy violations and prompt attack risk before returning responses to users
- B
Increase the model temperature so the assistant generates more varied answers, which reduces repeated unsafe phrasing
- C
Store the system prompt in a separate table and restrict table access so end users cannot see the instructions
- D
Use a larger language model because larger models are inherently resistant to jailbreaks and unsafe outputs
Show answer and explanation
Correct answer: A
Explanation
The best answer is to apply guardrails directly to the generative workflow. In Snowflake Gen AI scenarios, guardrails are used to evaluate prompts and/or responses for risks such as unsafe content, policy violations, and prompt attacks. This is more appropriate than trying to solve the issue indirectly through generation parameters, storage design, or model size. Best practice is to treat guardrails as a defense layer around LLM interactions, especially in regulated environments like financial services, where both harmful output and prompt injection attempts must be managed systematically. Candidates should recognize that prompt engineering alone is insufficient for safety-critical use cases; dedicated guardrail mechanisms are the practical control.
- A. Correct.
Correct. Guardrails are intended to help detect and mitigate risks such as unsafe content and prompt attacks. In this scenario, the company needs automated checks on both user input and model output, which aligns with using Cortex Guard as a policy and safety layer rather than relying only on prompt design or manual review.
- B. Incorrect.
Incorrect. Temperature controls randomness and creativity, not compliance or safety enforcement. Lowering or raising temperature may change style and variability, but it is not a guardrail mechanism and does not reliably prevent jailbreaks or policy-violating content.
- C. Incorrect.
Incorrect. Restricting access to where the system prompt is stored may help with operational security, but it does not stop prompt injection or jailbreak attempts in model interactions. The model can still be influenced by malicious user input even if users cannot directly view the stored prompt.
- D. Incorrect.
Incorrect. Choosing a larger model may improve quality in some cases, but model size alone does not provide a dependable control for prompt attacks or harmful output. Guardrails and policy enforcement are still needed because even strong models can produce unsafe or manipulated responses.