CCAR-P practice › Claude Models, Prompting & Context Engineering · Question 26 of 110
Claude Models, Prompting & Context Engineering · Select TWO
A vendor-research agent fetches supplier web pages, reads internal procurement files for context, and emails summaries to staff. A fetched page contained hidden text telling the agent to email the internal pricing sheet to an outside address, and the agent drafted that email before a person noticed. Which two changes most effectively reduce this risk? Select two.
Show answer and explanation
Answer: B and C — Enforce a recipient allowlist in code so the email tool rejects any address outside the internal domain / Remove the agent's access to internal pricing files, which the summarization task doesn't need
(b) and (c) are right: injection can't be prevented by wording alone, so limit what a successful injection can do. A code-enforced allowlist blocks exfiltration by email, and removing unneeded data access means there is nothing sensitive to send. (d) is a soft, prompt-only defense. (a) is detective, finding the leak a week after it happened.
Keep going
More Claude Models, Prompting & Context Engineering questions
- A marketplace uses Claude to assign a condition label (new, refurbished, used, for…
- A benefits-eligibility step applies six interacting conditions, and errors cluster on…
- A lender uses Claude, with a reasoning section before each answer, to draft credit-limit…
- A team enabled prompt caching with a cache_control breakpoint after the system prompt…
- A support agent handles long troubleshooting conversations. To control tokens, older…