Inference Controls
Ask models anything. Expose nothing.
Protegrity guardrails keep sensitive data out of prompts and block re-identified data in responses. NeMo, OpenAI and others plug in for behavior and accuracy.
Inference controls
Model interactionBefore the model
Input protectionAfter the model
Output protectionHow it works
Every prompt and response inspected.
Guardrails intercept prompts and responses between applications and models, run the checks configured for each application and return one outcome: allow, block, log or protect. Every request produces audit-ready telemetry.
Guardrail inspection
Application
Sends promptsReceives responses
Configured guardrails
Data
Sensitive values PII detectionProtection methods
Semantic
Intent and context Domain-specificmodels
Behavior
Allowed actions NVIDIA NeMoGuardrails AI
Model
Processes promptsGenerates responses
Policy decision
Data Guardrails
Protects sensitive values in prompts and blocks re-identified data in responses before it reaches users.
Semantic Guardrails
Blocks unsafe queries, off-topic use and data leakage in real time based on each application’s rules.
Behavior Guardrails
Defends against prompt injection and unsafe execution by limiting the actions a model can take.
Text to Analytics
Governed natural-language analytics.
A plain-language question becomes a structured query against approved data sources. Results are protected according to each user’s access, and original values appear only for authorized users.
A knowledge graph gives the model the metadata and relationships it needs to write the query, so sensitive values never reach the model. It works across data platforms.
User question
How many people in each department earn over $150K?
Query representation
SELECT dept, COUNT(*)
FROM employees
WHERE salary > 150000
GROUP BY dept;Salary values remain protected in the data source.
Governed execution
SELECT dept, COUNT(*)
FROM employees
WHERE id IN (1, 3, 4)
GROUP BY dept;Matching IDs supplied by a governed mapping.
Query result
Counts on protected data
| Department | Count |
|---|---|
| Engineering | 2 |
| Finance | 1 |
What you get
Models reasoning over protected data.
Sensitive data kept from models.
Prompts and responses are inspected inline, so PII, PHI and PCI are protected automatically, whatever users type into a prompt.
One control point for every guardrail.
Protegrity, NeMo, OpenAI and other guardrails run through a single point, and new agents and workflows simply connect to it.
Exposure based on clearance.
Authorized users see real values and everyone else sees tokens, even for data they submitted themselves.
Self-service analytics.
Business users ask questions in plain language and get answers scoped to their access, without writing SQL.
Audit-ready records.
Every request records what was checked and what each guardrail decided, exportable for the EU AI Act, SOC 2 and NIST AI RMF.
One flow from policy to enforcement.
Policy defines what to protect, guardrails enforce them inline, and clear values are resolved based on permissions.
Governed model interaction
Request and responseApplication context
AI Gateway
Policy-driven controlsProtection that works inside every request.
Protect and restore
Sensitive values are protected inline and restored per policy, so AI keeps working where most tools only detect or block.
Shared classification
Guardrails use Protegrity’s own discovery and classification, the foundation behind Data Protection and Agent Security.
Tokens survive the pipeline
Tokens hold up through chunking, embedding and transformation, so values stay resolvable downstream.
Protected AI interaction
Request and responseRequest origins
Configured guardrails
Applied according to application policyData protection services
Model
AI model Processes protected input and generates responsesProtect sensitive values according to policy.
Inspect outputs and restore values only when authorized.
Use cases
Protected going in and coming out.
Sensitive data in agent context
A support agent uses a tool to fetch a customer record, and the SSN lands in the prompt. It’s replaced with a token before the prompt reaches the model.
Agent context
Fetched customer record
- customer_id
- 8812
- ssn
- 078-05-1120
Model context
- customer_id
- 8812
- ssn token
- 412-93-7586
The SSN stays tokenized in the agent’s context.
Sensitive data in user prompts
An employee pastes their own date of birth into a prompt but lacks permission to see it unprotected, so it stays tokenized. Submitting data doesn’t grant access to it.
User prompt
Original prompt
Check which benefit plans I qualify for.
My date of birth is 04/12/84
Model input
Original value remains protected.
Re-identification in responses
A model combines low-risk fields to infer a patient’s diagnosis and identity. The response is inspected on the way back, and the inferred data is protected.
Response inspection
Model draft
This patient is likely Jane Doe and is being treated for Type 2 diabetes.
Governed response
This patient is likely Kay Lind and is being treated for Dx-4417.
Inferred identity and diagnosis are protected.