llm-mask: keep personal data out of cloud AI

The problem

Every request to a cloud AI model carries whatever is in the text: customer names, phone numbers, bank card numbers, passwords pasted along with a log. Once sent, it is out of your control.

Cleaning texts by hand does not scale. A simple “replace with stars” breaks the conversation: the model can no longer tell one customer from another, and the answer comes back full of gaps.

How it works

Names, email addresses, phone numbers, bank card numbers, internal network addresses and passwords are replaced with labels such as [[EMAIL_1]]. The same value gets the same label in every message, so the model can follow a long conversation.
The real values are put back into the answer, so your customer reads a normal reply.
Numbers, amounts and dates are left alone, so totals and calculations stay correct.
Passwords and keys are hidden and never put back. A private encryption key stops the request altogether.
If hiding fails, nothing is sent.
Names of your customers and companies come from your own lists; nothing is guessed.

Ways to use it

As a gateway. Your programs send their requests through llm-mask instead of straight to the AI service; nothing else in them changes.

Inside your own program. Two calls: one before sending a text, one on the answer.

On a single text. Clean a log or a document before pasting it into a chat with an AI model.

Trial mode first. Traffic goes out unchanged, while llm-mask counts what it would have hidden. You see the risk before you switch anything on.

Why it matters for security and data protection

Less personal data leaves the company and reaches third parties — a common requirement of data protection rules.
A record of what was found and how often, without the values themselves.
It runs on your own computer or server and keeps no keys of its own: the keys of your programs only pass through it.
It does not depend on the rest of GeneralBrain and works with any program that talks to an AI model.