# Secure anonymisation — the cloud model never sees plain text | IT-INTERDEV #anticloudllm > How the Schleuse detects credentials, customer names and identifiers locally, replaces them with stand-ins that keep the meaning and translates the cloud AI’s answer back. Short film and explanation. #anticloudllm Language: en · Source: https://it-interdev.de/en/schleuse/secure-anonymization.html · Alternate (de): https://it-interdev.de/schleuse/sichere-anonymisierung.html · Updated: 2026-10-05 [Schleuse](/en/schleuse.html) / Secure anonymisation \#anticloudllm Feature · cryptic shadow · 0.1 Alpha # The cloud model never sees **plain text**. Credentials, customer names and identifiers do not leave your computer. The Schleuse detects them **locally**, replaces them with **stand-ins that keep the meaning** and translates the cloud AI’s answer back. 65 seconds · with sound · made-up example · The film is narrated in German · Portrait version for your phone: [watch 9:16](/schleuse/media/film9_65s.mp4) 01 / Detect ## Read first — **on your side** Before a prompt leaves the house, the Schleuse reads the text with several detectors one after another. Everything runs on your own hardware; no cloud service is involved. The results are merged. Fixed patterns ### What has check digits IBANs, email addresses, metering-point numbers, tax numbers and credentials follow fixed formats and are found reliably. Name detector ### People, companies, places Open detection components (Presidio, GLiNER) find names in running text, in comments and in log lines. Local language model ### The second reader A language model on your own hardware reads the text once more and catches what patterns and the name detector miss. 02 / Replace ## Meaningful instead of **\[COMPANY_1\]** Instead of a bracket there is a general, related word: “Agricultural machinery business Alpha” instead of “Müller Landtechnik GmbH”. The meaning of the text stays readable, and the cloud model answers more usefully. Amounts, dates and quantities with units stay unchanged. Your original · stays with you\# Error for customer Müller Landtechnik GmbH \# IBAN DE89 3704 0044 0532 0130 00 \# Amount 12,480.00 EUR What the cloud model sees\# Error for customer Agricultural machinery business Alpha \# IBAN IBAN_1 \# Amount 12,480.00 EUR Made-up, simplified example. General terms reveal the type of the parties involved; with rare roles or sole traders this can be enough to identify someone. Then the coarser level is safer. 03 / Back and check ## Out and back must give **the original** The key between original and stand-in word stays with you. The cloud AI’s answer is translated back locally. A round-trip test checks that out and back gives exactly the original. If a stage fails or a remainder is left in the text, nothing goes out when in doubt. Round trip ### Exact or not at all If the way back differs from the original, the request is stopped and named in the report. Outgoing check ### Check before sending If a hit from the key remains in the outgoing text, nothing goes to the cloud model. Log ### Without contents The log records accesses and approvals, without contents, and is protected against later changes. 04 / Limits ## What we **say openly** ### Status: version 0.1 Alpha - So far tested only with made-up test data, mainly texts and documents. A separate test series for source code, configuration files and logs is still to come. - Not one hundred percent: detection is never complete. The remainder is measured and limited with thresholds; it cannot be ruled out. - Texts, documents, images and scans are handled. Audio and video are not yet. - No external review, no legal review. Our evidence is technical and not a legal certificate. Legally we treat the method as pseudonymisation as long as the key is with you. [Request a pilot](/en/#kontakt) [Back to the Schleuse overview](/en/schleuse.html)