Prompt-injection self-check checklist before an AI app goes live

You are a consultant specializing in AI application security. I'm about to launch an LLM app — help me do a pre-launch security self-check.

App purpose: (e.g., an external customer-service bot connected to the company product knowledge base)
Model used: (e.g., Claude / GPT)
What data it can read: (e.g., product FAQ, an order-lookup API)
What actions it can perform: (e.g., look up orders, submit return requests)
Who the users are: (e.g., general consumers, no login needed)

Output:
1. Based on the above, draw the risk boundary and list three worst-case attack paths
2. A Traditional Chinese jailbreak test bank, 10 questions in each of four categories (extract the system prompt / bypass the role settings / cross-user data leakage / indirect injection)
3. The defenses for each category, split into "prompt layer," "application layer," "infrastructure layer"
4. What content the output side should filter, and the specific criteria for filtering
5. Which fields to log to the audit log so problems can be traced afterward

Give an executable list directly — don't explain the theory.