A banking multi-agent assistant must never recommend specific securities trades or promise guaranteed returns, even if users jailbreak the base model safety filters. Compliance wants automated proof that custom policies hold. What should you implement?
Select an answer to reveal the explanation.
Short Explanation
You implement custom domain guardrails on inputs and outputs, then prove they hold with synthetic adversarial tests before each release. Nicer prompts alone aren't enough—you encode the banking policy and fire-drill it with jailbreak attempts.
Full Explanation
Domain-specific constraints require custom guardrails beyond default content safety, plus testing and validation including synthetic data. The correct approach is to implement custom input/output guardrails that enforce banking policies, such as no personalized trade recommendations and no guaranteed returns, then validate those rails with synthetic adversarial suites before each release.
The approach that increases temperature on the trading-advice agent is incorrect because higher temperature increases randomness and can worsen policy violations; it isn't a compliance control.
The approach that adds a longer system prompt only, without automated evaluation, is incorrect because system prompts help but are insufficient alone; the exam skill set explicitly includes guardrail testing with synthetic data.
The approach that trains a separate foundation model from scratch solely to refuse investment language, replacing hosted models entirely, is incorrect because training a foundation model from scratch is disproportionate, costly, and unnecessary when policy can be enforced with guardrails and evaluations on hosted models.
Exam tip: Custom domain guardrails plus synthetic validation is the pattern when industry policy exceeds generic safety filters.