Test one slice, but choose a slice that crosses the riskiest part of the problem. A tiny easy demo will only prove the easy part. Use this decision rule: - If reversing the choice is cheap, try it. - If the choice locks in data, money, or other people’s time, write the full decision first. - If uncertainty is concentrated in one step, test that step before planning around it. For prompt injection checks, make the first test produce evidence you can compare, not just something that looks complete.

This is usable as written. The important parts are the definition of done, the instruction to expose assumptions, and the request for a smallest test. For prompt injection checks, I would start with these checks: - Put the source of truth in the prompt - Mark missing evidence instead of filling it in - Add a stop condition before tool use If you share real material, remove anything private first and replace bracketed fields with concrete details.

Use the attached prompt by replacing its input bullets with the real context. The strongest results come from evidence-bearing inputs and explicit success criteria. The required deliverables are: architecture risk map, ranked findings with severity and evidence, remediation checklist, red-team test cases.

## Copy-ready prompt ```text Act as an application security engineer. Answer this question: “How do we threat-model prompt injection?” Objective: threat-model direct and indirect prompt injection in an AI application. First, ask up to five concise questions needed to tailor the analysis. If no answers are available, state reasonable assumptions and continue. Use current information as of today; browse authoritative primary sources where possible, distinguish facts from inference, and cite every time-sensitive claim with a direct link and publication date. Cover: trust boundaries, untrusted con