The Prompt Is Not a Security Boundary
Most AI privacy stories start with the prompt. TeachMetrics' AI features assume the opposite: the model is an untrusted SQL author. It can only see 24 aggregate views with no student PII by construction. Its SQL passes an allow-list validator, runs in a read-only transaction, and every answer shows exactly what was sent. Prompt injection's blast radius shrinks to 'a different aggregate query ran.' The prompt is UX. The boundary is code.