Prompt privacy
Prompts routinely contain customer names, emails, IDs and internal content. Without a security layer, all of that is sent to a third-party LLM.
Category · Educational pillar
What is an AI Security Layer?
Reference definition and architecture for the privacy-first LLM gateway category.
A vendor-neutral primer for engineering and security teams shipping AI features. For how Privian implements this layer, see the product page.
Definition
Definition
An AI Security Layer is the masking, routing and rehydration layer that sits between an application and one or more AI providers. It detects personal and sensitive data in inbound prompts, replaces it with deterministic placeholders, forwards the masked prompt to the configured provider, and rehydrates the response on the way back.
Why
Why AI systems need an AI Security Layer
LLM security
Direct provider calls leave no enforcement point for policy, masking, or auditing. The security layer centralizes that surface.
Sensitive data exposure
Even short-lived provider logs can become a compliance problem. Masking at the edge shrinks the exposure window to zero.
Customer data in prompts
Product AI features blend customer records with user input. The security layer strips identifiers before that mix leaves your stack.
Internal copilots
Employees paste internal documents into chat UIs. A gateway routes those prompts through masking before any external call.
Centralised policy point
A central choke point makes it tractable to apply structural validation and policy on every prompt — not scattered across clients.
How it works
How an AI Security Layer works
Step 1
Request routing
Your app POSTs prompt + model to the gateway endpoint instead of calling the provider directly.
Step 2
PII masking
Detected personal and sensitive values are replaced with stable placeholders inside the request body.
Step 3
Prompt security
Structural and policy checks run at the edge. Disallowed prompts are rejected with a structured error.
Step 4
LLM gateway routing
The masked prompt is forwarded to the configured provider using your own credentials (BYOK).
Step 5
Rehydration
Placeholders in the response are restored to original values before reaching your application.
Step 6
Zero retention
Raw prompts and responses are never persisted. Only structural counters are kept for observability.
Comparison
AI Security Layer vs LLM gateway
An LLM gateway is a routing layer for AI traffic — it accepts prompt requests, resolves the provider, and forwards the call. That alone does not change what the provider sees.
An AI Security Layer is the broader category. It wraps the gateway with PII masking, prompt security, rehydration, and a zero-retention guarantee. Every AI Security Layer contains a gateway; not every gateway is an AI Security Layer.
Capabilities
Core capabilities
Masking
PII masking
Deterministic detection and replacement of personal identifiers.
Policy
Prompt security
Structural validation and policy enforcement at the edge.
Privacy
Zero retention
Structural counters only — never raw prompt or response bodies.
Architecture
BYOK
Use your own provider credentials. Privian routes, not resells.
Routing
Provider-agnostic routing
One JSON gateway in front of multiple providers, selected via the model namespace.
Egress
Rehydration
Restore masked placeholders in the response so your app sees real data.
FAQ
Frequently asked questions
- What is an AI Security Layer?
- An AI Security Layer is a network hop between your application and AI providers (OpenAI, Anthropic, others) that masks personal and sensitive data, enforces prompt security policies, routes requests using your own provider keys, and rehydrates the response — so the model never sees the original values.
- How is an AI Security Layer different from an LLM gateway?
- An LLM gateway is the routing component. An AI Security Layer is the broader category: an LLM gateway plus PII masking, prompt security, rehydration and zero-retention guarantees. Every AI Security Layer contains a gateway; not every gateway is an AI Security Layer.
- Why do AI systems need an AI Security Layer?
- Customer-facing AI features and internal copilots routinely send personal data — names, emails, customer notes, internal documents — into third-party LLMs. An AI Security Layer prevents that exposure by masking sensitive values before they leave your stack.
- Does an AI Security Layer store prompts?
- Privian does not. Raw prompts and responses are never persisted. Only structural counters (token counts, latency, masked entity types) are kept for observability.
- Does it work with my existing OpenAI or Anthropic account?
- Yes. Privian uses a BYOK (bring-your-own-key) model. Provider credentials are encrypted at rest and decrypted in-process at request time. Usage continues to bill against your own provider account.