Explanation:
Part 1 (Protocols/Ciphers):
API Gateway: TLS 1.2
* Database: AES-512
* AI Model: gRPC
Part 2 (Techniques):
1. {pin:"999-99-9999", name:"john doe"} → {pin:"999-99-9999", name:"john doe", sensitivity:"SECRET"} → Classification
2. {ip_addr:"1.2.3.4", cookie:"aK3idkd==", name:"John Doe" uid="1111"} → {cookie:"aK3idkd==", uid="1111"} → De-identification
3. {pin:"999-99-9999", name:"john doe"} → {pin:"999-99-XXXX", name:"john doe"} → Masking
4. {card_number:"1111222233334444", name:"john doe", user_id:"1"} →
{card_number:"0x019238829", name:"john doe", user_id:"1"} → Tokenization
5. {name:"john doe", patient_id:"10000", dob:"1980-Jan-05"} → {patient_id:"10000", dob:"1980- Jan-05"} → Anonymization Part 1:
* TLS 1.2 secures client-to-gateway communications.
* AES-512 provides strong encryption for data at rest in the database.
* gRPC ensures efficient, secure communication between services (AI model).
Part 2:
* Classification tags sensitive data for handling.
* De-identification strips direct identifiers.
* Masking obscures part of sensitive values while keeping format.
* Tokenization replaces sensitive data with a reversible placeholder.
* Anonymization removes identifying attributes, making re-identification impossible.

Providing examples in a prompt is referred to as one-shot prompting when a single example is given and multi-shot prompting when multiple examples are provided.

Basic Concept: LLM API pricing is primarily based on token consumption - the number of tokens processed in both input prompts and output responses. Controlling token usage is the most direct lever for managing and reducing LLM API costs. CompTIA SecAI+ Study Guide covers AI cost management and resource controls under securing AI systems.
Why D is Correct: Adjusting token limits directly caps the maximum number of tokens used per request for both input and output. By setting appropriate token limits, the organization prevents excessively long prompts or verbose responses from consuming unnecessary tokens, directly translating to lower API costs and providing hard budget control.
Why A is Wrong: Prompt templates standardize how queries are structured, which can indirectly improve efficiency. However, they do not enforce a hard cap on token usage and cannot prevent costs from escalating with large volumes or verbose responses.
Why B is Wrong: Increasing CPU and memory addresses computational infrastructure performance on the client side. LLM API costs are billed by the API provider based on token usage, not on the client ' s hardware resources.
Why C is Wrong: Reducing model size means using a smaller, less powerful model version. While this may lower cost per token, it is a model selection decision, not an ongoing operational control that can be adjusted to manage cost in real time.

Option B is correct because output filtering examines the chatbot's generated response before it reaches the requester and blocks or redacts sensitive information such as a patient's phone number. In this scenario, the requester uses urgency and a claimed family relationship to pressure the model, but the chatbot must not disclose protected personal data without verified authorization. An output filter can detect personally identifiable or health-related information, apply policy rules, replace the value with a refusal, and log the event for review. Option A standardizes prompts but cannot guarantee that the model will not produce sensitive content. Option C encrypts the communication channel, protecting data from interception in transit, but it would still deliver the prohibited information securely to an unauthorized person. Option D masks data before model use and can reduce exposure, but the question asks for a compensating control that prevents disclosure in the chatbot response; runtime output filtering is the most direct choice. The NIST AI Risk Management Framework links trustworthy AI with privacy enhancement and protection of confidentiality, supporting controls that prevent inappropriate disclosure.

Guardrails are the primary security control for LLMs to prevent prompt injection attacks. They enforce rules on what inputs are accepted and how the model responds, blocking malicious or sensitive prompts (such as credit card numbers) before they can manipulate or exploit the model.