FASTEERMASK replaces sensitive data with labels that tell the model what type of data it is, not which data it is. The model still responds well. Real data never leaves your perimeter.
Works with the models you already use. No migration, no retraining.
The label preserves the meaning of the data, so the model can continue to reason correctly.
Clear text never crosses the boundary to the model provider.
The data-to-label mapping lives in volatile memory and is deleted after use.
Whenever someone pastes a contract, medical record, or customer list into an external model, that data leaves the company. Existing solutions share the same flaw: they protect the data by weakening the answer.
Names, tax IDs, IBANs, health data, project names, and product names end up in the prompt and leave the company perimeter.
Remove the data before sending and the model loses context, producing lower-quality answers.
Replacing data with fake values does not solve the problem: the model no longer understands what kind of information it is seeing.
The example below is taken from the patent text. The values are fictional; the mechanism is the one running in production.
| Detected data | Assigned label |
|---|---|
| Mario Rossi | <NAME 45> |
| Milano | <CITY 37> |
| January 1, 1990 | <DATE 104> |
| PRJ_0 | <PROJECT 12> |
The mapping stays in volatile memory inside your perimeter. The number next to a name distinguishes different people in the same text.
The model knows that <NAME 45> is a person and <DATE 104> is a date. It does not know who or when.
The labels become real data again by reading the local mapping. The mapping is then deleted. The user reads a complete answer; the model provider has never seen real data.
Redacting text is easy. Redacting it while leaving the model enough context to answer well is the patented idea.
Not a random code: a name semantically related to the replaced data, so the model knows it is handling a person, an address, or an IBAN.
A unique identifier distinguishes different people in the same text, so the model does not confuse them.
Real data is restored locally. There is no moment when the model provider can see it.
The data-to-label mapping is temporary and deleted after use: no second archive accumulates that needs protection.
Remove anything that does not need protection and add anything automatic detection missed.
The same mechanism protects project names, product names, and confidential concepts: trade secrets, not just personal privacy.
Regular expressions for structured formats, named-entity recognition for people and organizations, and classification for sensitive categories.
Anywhere people are already pasting company documents into an external model, with or without authorization.
Analyze and summarize contracts without exposing counterparties, amounts, or confidential clauses.
Screen CVs and personnel documents without exposing identifying data.
Report on and summarize clinical text while keeping the patient unidentifiable.
Process case files with IBANs, cards, and identity data protected at the point of entry.
Draft ticket responses without sending customer data outside.
Project names, unannounced products, and internal concepts stay protected.
FASTEERMASK is a layer between your applications and the model. It does not replace the model or change your stack.
Works with leading commercial models and open models installed in-house.
Fits between the applications and agents already querying models, without rewriting them.
European cloud or installation in your infrastructure, with data and memory inside your boundary.
Available as a module of the FASTEER platform, activated inside the customer instance.
Decide which categories to always mask, leave to the user, or ignore.
What was masked and when remains recorded: the evidence you need when someone asks for an audit trail.
This is not a badge: it is the fact that data does not leave, documented and demonstrable.
Minimization applied at the right point: personal data is not disclosed to the model provider.
A verifiable technical control over the data feeding the AI systems you use.
Project names, products, and confidential concepts protected in the same way as personal data.
This is not a feature built on top of a model: it is a registered method covering the full path of masking, model interaction, and reconstruction. The claims explicitly cite Large Language Models.
FASTEERMASK is a product based on AidaMask technology by Aida S.r.l., a company of the FoolFarm S.p.A. group. The patented technology masks sensitive data before it reaches external models. Developed in Italy, it is protected by Italian industrial invention patent no. 102023000007785, owned by FoolFarm S.p.A.
That is the point: no. The label tells the model what type of data it is, so reasoning stays correct. That is exactly what distinguishes this method from deletion or random substitution.
It stays where it is. The data-to-label mapping lives in volatile memory inside your perimeter and is deleted after use.
No. FASTEERMASK sits between your applications and the model you already use, commercial or open, in the cloud or installed in-house.
The detected-data list can be edited before sending: remove what does not need protection and add what was missed.
No. The same mechanism covers project names, product names, and internal concepts, which are the most concrete risk for many companies.
As a module of the FASTEER platform, or as a component integrated into your architecture. We start with one of your use cases and show it on your documents.
Tell us what your teams work with today. We will show the same question with and without masking, using your material instead of a slide.