What enters the model. What leaves. Both controlled.
AI models process exactly what they receive. Without interception, PII and PHI reach external models and sensitive institutional data appears in responses. Privacy Sentinel intercepts both directions structurally — before any data crosses the model boundary.
Inbound: the prompt is inspected before reaching the model. Personal identifiers, health data, financial references, and sensitive categories are detected and handled per operator policy. Outbound: the model's response is scanned before delivery to the user. Data leakage and out-of-scope content are intercepted at the infrastructure layer — not in application code that can be bypassed.
Prompt injection attempts, social engineering patterns, and adversarial instructions are identified and blocked before reaching the model. The manipulation gate at Gate 06 operates as a dedicated enforcement layer — independent of the privacy gates at 05 and 08 — covering distinct threat vectors with separate detection logic.
Every prompt is analyzed for PII, PHI, financial identifiers, and sensitive categories before reaching any model. Detection is structural and cannot be bypassed by application code.
Model responses are scanned before delivery. Data leakage, policy violations, and out-of-scope content are intercepted at the infrastructure boundary, not after delivery.
Prompt injection, social engineering, and adversarial instruction patterns are blocked at Gate 06 — a dedicated enforcement layer separate from the privacy inspection gates.
The enforcement response is operator-configured per data category: automatic redaction, request blocking, or flagging for human review. All three options produce a signed audit record.
Technical details are disclosed only within controlled institutional engagements.