The National Institute of Standards and Technology entrance sign at its Gaithersburg campus
AI privacy“What did your AI see?” should have a technical answer.
NIST's Gaithersburg campus entrance. NIST guidance is referenced in this article.Image: R. Wilson / NISTNIST reuse policyImage fitted for layout; headline added by Securelay. Editorial context, not endorsement.
Primary source NIST AI 600-1 — Generative AI ProfilePrimary source EDPB Opinion 28/2024 on AI models and personal data

Begin at the data path

Before a prompt, retrieved context, or tool result reaches a model connection, detect configured sensitive classes and replace them with stable references. The model receives enough structure to perform the task without automatically receiving the source identifiers.

Rehydration is disclosure

Putting a sensitive value back into a response is not a formatting step. It is a disclosure decision. It should evaluate current role, purpose, consent, scope, and operational conditions, and it should be auditable whether the answer is permit or deny.

Manifest the session

A session manifest can link provider configuration, transformation events, governed disclosures, subject and purpose context, and evidence handles. This gives investigators a bounded account of the protected path.

Do not overclaim the boundary

A privacy gateway does not control an agent that sends traffic around it. It does not choose the model, sandbox tools, solve prompt injection, or shut down the orchestration runtime. Network egress, runtime permissions, and agent controls remain essential companion controls.

Put the control on the data path.Discuss an architecture review