SiperBook a demo
Menu

Personal data masking

Siper masks personal data in Turkish text before it reaches an AI model.

Siper replaces each value it finds with a tag such as <KİŞİ_1>, so the model still reads a whole sentence.

On health data, Siper finds 81% and the next system finds 51%. Across all personal data, it finds 85.7% and the next system finds 69.3%.

Measured 2 November 2026, "Turkish and English" option. The model is trained for Turkish, and it detects from context, not only from patterns.

Sample

Müşteri Selin Aksoy depresyon tedavisi gördüğü için kredi taksitini ertelemek istiyor.

"Customer Selin Aksoy wants to defer her loan instalment because she is being treated for depression."

AI receives

Müşteri <KİŞİ_1> <SAĞLIK_1> için kredi taksitini ertelemek istiyor.

What it finds

Siper finds sensitive data in plain sentences, not only in ID numbers.

A national ID is a number pattern, and many tools find it. A diagnosis written inside a sentence is a different job. Siper reads the phrase in its Turkish context, for example "diyabet tedavim" ("my diabetes treatment") or "Örnek-İş Sendikası üyesidir" ("is a member of the Örnek-İş union").

Name, diagnosis and criminal record in one sentence

Müvekkilim Nurten Aksay'ın epilepsi tanısı ve adli sicil kaydı, işten çıkarma dosyasına gerekçe olarak yazılmış.

"My client Nurten Aksay's epilepsy diagnosis and criminal record were written into the dismissal file as grounds."

AI receives

Müvekkilim <KİŞİ_1>'ın <SAĞLIK_1> ve <CEZA_1>, işten çıkarma dosyasına gerekçe olarak yazılmış.

The name, the diagnosis and the record became tags.

  • Identity and contact

    • Person name
    • Address
    • Email
    • Mobile phone
    • Landline
    • Date of birth
    • Turkish national ID (TCKN)
    • Foreign resident ID (YKN)
    • Foreign country ID
    • Passport
    • Driving licence
    • Bar registration number
  • Financial and technical

    • IBAN
    • Card
    • Tax ID (VKN)
    • Company registry number (MERSİS)
    • Social security number (SGK)
    • Plate
    • Serial number
    • IP address
    • MAC address
    • Secrets: API key, JWT, private key

Siper covers 32 data categories. The lists above give examples, and seven classes of KVKK Article 6 special-category data are in the next section.

Same text, two methods

A pattern list masks 1 value in this text. Siper masks 5 values.

Both methods read the same synthetic Turkish text. This pattern list has five patterns: national ID, IBAN, mobile phone, email and IP address.

Input text

Teknisyenimiz Mehmet Yılmaz (TCKN 10000000146) için görev yeri değişikliği yazısı hazırla. Epilepsi tanısı nedeniyle arazi görevine verilmemesini ve sendika temsilcisi olduğunu belirt. Taksirle yaralamadan aldığı ertelenmiş cezayı yazıya koyma.

"Prepare a letter on the change of work location for our technician Mehmet Yılmaz (national ID 10000000146). State that he should not get field duty because of his epilepsy diagnosis, and that he is a union representative. Leave out of the letter the suspended sentence he received for negligent injury."

  • A pattern list

    1 value masked

    AI receives

    Teknisyenimiz Mehmet Yılmaz (TCKN <TCKN_1>) için görev yeri değişikliği yazısı hazırla. Epilepsi tanısı nedeniyle arazi görevine verilmemesini ve sendika temsilcisi olduğunu belirt. Taksirle yaralamadan aldığı ertelenmiş cezayı yazıya koyma.

    It masks the national ID. The name, the diagnosis, the union role and the criminal record stay in the text.

  • Siper

    5 values masked

    AI receives

    Teknisyenimiz <KİŞİ_1> (TCKN <TCKN_1>) için görev yeri değişikliği yazısı hazırla. <SAĞLIK_1> nedeniyle arazi görevine verilmemesini ve <SENDİKA_1> belirt. <CEZA_1> yazıya koyma.

    It masks the name, the national ID, the diagnosis, the union role and the criminal record. The rest of the text stays the same.

One example, not a measurement: one text, with the Siper result recorded from the public demo on 5 Nov 2026. Read the benchmark for the measured result.

Language options

You choose one of three language options on each request.

  • The options are Turkish, English, and Turkish and English. The price is the same for all three.
  • Siper masks values it recognises by format under every option: Turkish national ID, IBAN, phone number and email address.
  • Under the English option, special-category data in Turkish text is caught less often.
  • The numbers on this site use the "Turkish and English" option, on Turkish text.

KVKK Article 6

Siper masks seven classes of KVKK Article 6 special-category data.

Article 6 of Law 6698, the Turkish personal data law, defines special-category personal data. KVKK's page on these data repeats the list, and it is longer than this table. Read 5 Nov 2026.

Siper recall for each KVKK Article 6 category
CategorySiper recall
Health81%
Religion86%
Ethnic origin91%
Political opinion88%
Union membership94%
Criminal conviction77%
Sex life90%

Measured on 403 human-verified Turkish documents, October 2026.

Read the benchmark

Reversible masking

The vault can restore real values in the answer, and the model still receives the masked text.

By default, the answer comes back with the tags as they went out. If you turn on the vault, Siper restores the real values before the answer reaches you.

What the model and your application see with the vault off and on
SettingThe model seesYour app sees in the answerWhat is kept
Vault off (default)<KİŞİ_1><KİŞİ_1>No mapping from tags to values is described for this setting.
Vault on<KİŞİ_1>Selin AksoyEach tag and its original value, in a vault.

The direct masking API returns English tag names such as <PERSON_1>. The gateway and the panel use Turkish tags. The vault applies to model answers.

Before and after

The sentence stays readable, and the values it found become tags.

Each sample shows the original text, then the text the AI model receives.

Customer complaint

Merhaba, ben Ayşe Karahan. 12 Mart'ta Kadıköy şubenizde yaptığım başvuruya hâlâ dönüş yapılmadı. Bana 0532 000 00 00 numaralı hattımdan ya da ayse.karahan@example.com adresinden ulaşabilirsiniz. TC kimlik numaram 10000000214. Sağlık raporumda belirtilen diyabet tedavim nedeniyle sürecin hızlandırılmasını rica ediyorum.

"Hello, I am Ayşe Karahan. My application at your Kadıköy branch on 12 March has still had no reply. You can reach me on 0532 000 00 00 or at ayse.karahan@example.com. My national ID number is 10000000214. I ask you to speed up the process because of the diabetes treatment stated in my health report."

AI receives

Merhaba, ben <KİŞİ_1>. 12 Mart'ta <ADRES_1> şubenizde yaptığım başvuruya hâlâ dönüş yapılmadı. Bana <TELEFON_1> numaralı hattımdan ya da <EPOSTA_1> adresinden ulaşabilirsiniz. TC kimlik numaram <TCKN_1>. Sağlık raporumda belirtilen <SAĞLIK_1> nedeniyle sürecin hızlandırılmasını rica ediyorum.

Bank application

Müşteri Elif Korkmaz, TC 10000000214 ile başvurdu.

"Customer Elif Korkmaz applied with national ID 10000000214."

AI receives

Müşteri <KİŞİ_1>, TC <TCKN_1> ile başvurdu.

Patient record

Hasta Mehmet Yıldız, TC 10000000382 ve diyabet tanısıyla izleniyor.

"Patient Mehmet Yıldız, national ID 10000000382, is being monitored with a diabetes diagnosis."

AI receives

Hasta <KİŞİ_1>, TC <TCKN_1> ve <SAĞLIK_1> tanısıyla izleniyor.

Where it applies

Masking applies to employees, applications and agents, on different paths.

Entry points for masking and what Siper does at each
Entry pointWhat Siper does
EmployeesEnforces your rules on each request: it masks personal data as requests go out to ChatGPT, Claude, Gemini, Copilot, Perplexity, DeepSeek, Mistral and Grok. It also covers Word, Excel, PowerPoint, PDF, CSV and images, with scanned pages through OCR.
ApplicationsMasks text through the masking API, or inside the AI gateway when you change the base URL and API key.
AgentsClaude Code and Codex run through the Siper gateway. Data returned by MCP tools is masked before it reaches the model. The public MCP demo offers one masking tool to clients such as Claude and Cursor.

KVKK, the Turkish data protection authority, published a notice on workplace generative AI tools in March 2026. Its document says general wording is more cautious than details such as specific names, dates and locations. Masked data stays personal data. Have counsel review Article 9. KVKK says pseudonymised data stays personal data, in a notice about call logs, not AI. Read 5 Nov 2026.

See how each path installs

Next step

Test it on your own Turkish text.

Choose a short sample from your own work. In a demo we check together what Siper finds.

Book a demoTry the playground

Use made-up data, not real personal data.