Which defense-in-depth approach BEST reduces prompt injection risk for a Claude app that reads untrusted documents?
Select an answer to reveal the explanation.
Short Explanation and Infographic
Untrusted docs are data, not bosses. Sandbox them, limit tools, and validate every tool call on your server.
Full explanation below image
Full Explanation
Prompt injection exploits models treating untrusted content as instructions. Mitigations: clear separation of system vs data, least-privilege tools, server-side authorization of tool effects, output filtering, and human approval for high-impact actions.