BlueHomer
Protecting data in the age of public AI
BlueHomer treats external AI models as swappable compute, never as owners of your data. Between your enterprise and any model sits one mandatory gateway that cannot be disabled.
The guarantee
When your data is sent to an external, public AI model, it is protected first — sensitive values are tokenized, redacted or blocked according to your account’s policy. Private, on-premise models receive real data. You choose the trade-off; protection on the public path is not optional.
What happens on the way out
- Policy is evaluated first — before data is retrieved and before anything is sent. If the request is not permitted, it never leaves.
- PII detection and redaction — sensitive values are found and replaced before anything crosses the boundary.
- Tokenization and obfuscation — sensitive identifiers and schema are mapped to type-preserving tokens while the relationships between them are preserved, so the model can still reason correctly.
- Prompt-injection defense — inputs are inspected for manipulation attempts before they reach the model.
What happens on the way back
- Restoration — tokens are mapped back to their real values, so the user sees real, correct data that the external model never saw.
- Output scanning — responses are checked for leaked secrets and manipulated content.
- Full audit logging — every interaction is recorded for compliance.
This is the difference between redaction and tokenization in one line: redaction protects the value and destroys the relationship; type-preserving tokenization protects the value and keeps the relationship, then gives the value back on the response.
You choose the trade-off
| Route | What the model receives | What you get |
|---|---|---|
| Public frontier model | Tokenized, redacted or blocked values per your policy | Frontier capability with protection by construction |
| Private / on-premise model | Real data, inside your boundary | Full fidelity, no external exposure |
| Blocked by policy | Nothing — the call does not happen | A fail-closed default rather than a silent exception |
Either way, third-party models stay what BlueHomer intends them to be: swappable compute, never owners of your intelligence. Switching model providers does not change your data protection posture, because the posture lives in the gateway, not in the vendor.
Why it is mandatory
A protective control with an off switch is a control that will eventually be off. The gateway is the single front door to every model on the platform: there is no second path, no direct SDK call, no per-team exception. That is what makes the guarantee above a property of the architecture rather than a promise about behaviour.
Frequently asked questions
Is it safe to connect ChatGPT or another public model to company data?
Not directly. Pasting or piping enterprise data into a public model exposes personal data, schema and business context, and produces answers no one can audit. It becomes safe when a mandatory gateway sits in between — detecting and redacting personal data, tokenizing sensitive values with their relationships preserved, defending against prompt injection, scanning responses, and logging every interaction.
How is tokenization different from redaction?
Redaction removes the value — the model can no longer reason about it. Type-preserving tokenization replaces the value with a token of the same shape while preserving relationships between records, so the model still reasons correctly, and the real values are restored in the response.
What does the external model actually see?
Only what your policy permits: tokenized or redacted content, with sensitive identifiers and schema obscured. The user sees the restored, correct result.
Can we bring our own model?
Yes. Public and private models are interchangeable behind the gateway, which is what makes the platform vendor-agnostic at the AI layer.