ACCESS
Who decides what users can do?
Enforce every data operation on the server or trusted data layer, and test for privilege escalation.
SECURITY & QUALITY
What security checks does an AI-built app need?
Code generation can accelerate a prototype. Production readiness still requires explicit controls and independent tests.
Understand the work before you invest in it.
A PROTOTYPE IS NOT A CONTROLLED SERVICE
Verify the boundaries before production.
AI-assisted coding tools can assemble interfaces and connect services quickly, but generation is not proof that permissions, input handling and secrets management are secure. Application owners need independent review of what the system can read, modify and send.
Start with a system diagram: browser, API, identity provider, data store, third-party AI services and any customer-specific integrations. Annotate where personal information crosses a trust boundary and which component authorises each action.
Test with users who have different roles. A user should not be able to read, update or infer another customer's records by changing an identifier in a request. Server-side permissions must not depend on hiding a button in the interface.
TESTABLE CONTROLS
Questions the implementation must answer
ISOLATION
Can one customer reach another?
Validate cross-tenant isolation across database queries, search indexes, caches, files and background jobs.
OBSERVABILITY
How would a failure be discovered?
Record appropriate security events, track error rates and document who handles incidents.
RECOVERY
What is the restore procedure?
Check backup integrity, practise restoration and document realistic recovery objectives.
What about AI-specific risks?
If the app uses a language model, treat retrieved documents, user input and external tool outputs as untrusted. Apply explicit permissions to actions and tools; do not interpret model-generated text as an authorisation decision. Evaluate factual accuracy, prompt injection resilience and unintended data disclosure.
The exact controls depend on the application's risk profile. A low-risk internal checklist and a payment-handling or medical application should not share the same acceptance criteria.
For upstream product definition see MVP requirements. For data location and supplier questions see EU data residency explained.