SiperBook a demo
Menu

The AI firewall that runs inside your network.

Inspects prompts, files and agent calls on their way to ChatGPT, Claude or any other AI. Masks the personal data it detects, blocks prompt attacks and logs each decision.

Book a demoTry the playground

Your network

  • EmployeesPrompts and file uploads
  • ApplicationsChatbots, copilots, AI apps
  • AgentsCoding agents and MCP tools
  • Employees go through your web proxy, over ICAP. Nothing is installed on laptops.
  • Applications and agents call Siper's API. You point the client at a new base URL.

Siper

  • Masks personal data
  • Blocks prompt attacks
  • Logs each decision

Masked text

AI services

  • ChatGPT
  • Claude
  • Gemini
  • Copilot
  • Perplexity
  • DeepSeek
  • Mistral
  • Grok

Before Siper After Siper: masked text

Siper masks the personal data it detects, and the AI service receives the masked text. On our Turkish test set of 403 documents, Siper finds 85.7% of the personal data.

  • 85.7%

    recall on Turkish text

    Nearest alternative: 69.3%. Measured 2 November 2026, "Turkish and English" option.

  • On-prem

    runs inside your network

    or fully air-gapped

  • In-house

    AI models

    trained on our own datasets

  • No GPU

    Siper runs on CPU

    Measured latency is on the benchmark page.

Recorded in ChatGPT

A Word file and a scanned payslip go to ChatGPT. Siper masks both on the way out.

An employee drops a court petition and a scanned payslip into ChatGPT and asks for the personal data in them. Siper masks both at your web proxy: ChatGPT gets placeholders in the petition and black bars on the scan. Nothing is installed on the laptop.

The ChatGPT screens were recorded through Siper's web proxy path. The marks, the captions and the sound were added. The people and the data are made up.

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.

Proof

On health data in Turkish text, Siper finds 81%. The nearest alternative finds 51%.

Share of data found on the same Turkish documents: Siper against the nearest alternative in each row.
DataSiperNearest alternative
Health data81%51%
Criminal conviction data77%53%
All personal data (recall)85.7%69.3%

Siper covers 32 data categories, including seven classes of KVKK Article 6 special-category data. Measured 2 November 2026, "Turkish and English" option. We compared systems on 403 human-verified Turkish documents in October 2026.

Read the full benchmark

Who it protects

One firewall for your employees, your applications and your agents.

  • Usage control

    Employees

    Staff keep using ChatGPT, Claude and the other AI services. Siper masks the personal data it detects in their prompts and uploaded files at your web proxy.

    Employees: what Siper does
  • Application security

    Applications

    Point your app at a new base URL. Siper masks personal data and checks for prompt injection before the request reaches the model.

    Applications: what Siper does
  • Coding agents and MCP tools

    Agents

    Claude Code and Codex run through the Siper gateway. Siper masks data returned by MCP tools before it reaches the model.

    Agents: what Siper does

Connect

Connect in minutes, with what you already run.

Choose how to connect

Point your client at Siper by changing its base URL. The gateway is OpenAI-compatible and also serves the Anthropic Messages API.

Python, OpenAI SDK

import os
from openai import OpenAI

client = OpenAI(
    base_url="https://siper.example.local/v1",
    api_key=os.environ["SIPER_API_KEY"],
)

Your web proxy decrypts AI traffic and hands it to Siper over ICAP, so there is nothing to install on employee laptops.

Squid, the ICAP lines in squid.conf

acl siper_ai dstdomain "/etc/squid/siper/ai-domains.acl"
icap_enable on
icap_service siper1 reqmod_precache bypass=off icap://<siper server>:1344/siper
adaptation_access siper1 allow siper_ai
http_upgrade_request_protocols websocket deny all

Route a coding agent's model traffic and its tools through Siper. Tool results are masked before they reach the model. Set every model name your agent uses.

Claude Code

export ANTHROPIC_BASE_URL="https://siper.example.local/anthropic"
export ANTHROPIC_AUTH_TOKEN="$SIPER_API_KEY"
export ANTHROPIC_MODEL="<model id>"
claude mcp add --transport http siper https://siper.example.local/mcp \
  --header "Authorization: Bearer $SIPER_API_KEY"

See how Siper installs

Audit trail

Siper keeps a record of what it masked, without a copy of the text.

The panel keeps a summary: how many characters Siper processed, which categories it caught, and what it masked in which request. It does not keep the raw text.

This matters for staff use. KVKK, the Turkish data protection authority, warns employers against general review of work channel content.

Read about the audit trail

Sample audit rows: one row for each request, with the service, the content and the result
ServiceContentResult
ChatGPTpromptMasked
Claudereport.xlsxMasked
Agentinvoice.pdfBlocked

Sample rows, with made-up data.

Questions

What security teams ask first.

We already have a firewall and DLP. Why add Siper?

Keep them. A pattern finds the 11-digit Turkish ID. Finding "epilepsi tanısı" ("epilepsy diagnosis") inside a sentence is a different job. A tool that blocks also stops the work, and Siper masks, so the work continues. Run your own text through both and compare.

We have an enterprise AI contract. Is that enough?

A contract does not remove the transfer question. The large cloud providers we checked have no open Turkey region, so data sent to the large AI services goes abroad for now. With masking, the values that Siper detects stay inside your network.

We run our own LLM on our own servers. Do we need Siper?

Siper can route to your own model server, and it masks detected personal data before the request reaches the model.

Pilot

Free pilot: your own text, one Linux server, no GPU.

The pilot shows what Siper catches in your own text, what it misses and what it costs.

  1. First

    Connect one team or one app. Collect a sample set of your own text.

  2. Then

    See real numbers in the panel: how many requests, how many held personal data, and which categories.

  3. Result report

    The report says what was caught, what was missed, and the cost.

Next step

See what Siper finds in your own Turkish text.

Book a demo to test Siper on a short sample of your own text.

Book a demoTry the playground