Experimental, in testing
Pseudonymise your data before the ChatGPT, Claude or Gemini APIs
Your teams want to use the best models on the market, but those APIs send your data to third-party vendors. We are currently developing and testing an pseudonymisation layer that masks sensitive data before it is sent, then restores it when the answer comes back, so as to limit leaks while keeping the power of these models.
Status: experimental, in testing, currently in the testing phase. The elements presented below describe the approach we are validating before opening it up in production.
The problem
The problem: every prompt is a potential leak
As soon as a user copies a document into ChatGPT, Claude or Gemini, the data leaves your perimeter: surname, first name, email address, IP address, case file number, health data and more. That information travels and can be retained by the vendor. For many organisations this is unacceptable, but banning these tools only pushes people towards shadow IT.
The approach
The approach we are testing: mask, send, restore
Detection of sensitive data
An engine automatically spots personal data in the text: names, emails, IP addresses, numbers, addresses and more.
Replacement with tokens
Detected data is replaced with tokens. Undetected identifiers and identifying context may remain.
Sending to the third-party API
Only the pseudonymised text is transmitted to ChatGPT, Claude or Gemini. The mapping table stays with you.
Restoration on receipt
The response is restored locally using the mapping table. Masking does not guarantee that the transmitted text contains no personal data.
The round trip, in one diagram
Only the masked text crosses the boundary. The mapping table, used to restore values, must remain within your infrastructure with restricted access. Context may nevertheless allow reidentification without that table.
Example
A concrete example
Original text (stays with you)
“Please get back to Jean Dupont at j.dupont@ville.fr about case file no. 74123.”
Sent to the API (pseudonymised)
“Please get back to [PERSON_1] at [EMAIL_1] about case file [REF_1].”
Restored answer (user side)
“A follow-up email has been drafted for Jean Dupont (j.dupont@ville.fr), case file no. 74123.”
Limits
Defence in depth, not a miracle cure
Pseudonymisation sharply reduces the exposure of personal data, but it does not replace a sovereign AI: detection is never 100 % perfect, and the channel is still that of a third-party vendor. It is an excellent complement when the use of external APIs is unavoidable.
For full control, the best answer remains sovereign AI hosted in house, where the question of leakage no longer arises.
Testing phase
Take part in the testing phase
This solution is experimental (testing phase under way). Contact us to be told when it opens, or to take part in the trials with your own use cases.
References
Data protection: scope and status
The pseudonymisation gateway for external APIs is experimental and in testing; it is not advertised as production-ready. The mapping table allows restoration: data remains personal and GDPR still applies. Test detection and re-identification on the actual corpus.