Pricing Contact
Even the operator cannot read your data

Zero-Knowledge
Storage.

KI-Shield detects personal data automatically using ML models, replaces them with pseudonyms and encrypts everything with AES-256-GCM. What is stored cannot be read by anyone – not even us.

API + Web Chat • 42 PII categories • 6 AI providers • Made in Germany

How It Works

Three Protection Layers for Your Data

Every message automatically passes through three stages – without your team having to do anything.

Layer 1

ML Detection

Our language model detects 42 categories of personal data – names, IBAN, diagnoses, addresses, phone numbers and more.

Layer 2

Pseudonymisierung

"Max Mueller" becomes PERSON_001, an IBAN becomes IBAN_001. The AI provider never sees real data – but responds contextually correct.

Layer 3

AES-256-GCM

Everything stored – messages, mapping tables, titles – is encrypted with AES-256-GCM. Key derived from your password (Argon2id).

ki-shield ~ dataflow
# Your message:
> „Herr Müller (IBAN DE89 3704 0044) hat Diabetes Typ 2.“
 
━━━ ML Detection + Pseudonymization ━━━
Detected: Müller → PERSON_001 • DE89...0044 → IBAN_001 • Diabetes Typ 2 → HEALTH_001
 
━━━ What the AI Provider Sees ━━━
> „Herr PERSON_001 (IBAN IBAN_001) hat HEALTH_001.“
 
━━━ What Is Stored in the Database ━━━
Message: gcm$aG9yc2V...kQ==
Mapping: gcm$bXkgc2V...wA==
Title: gcm$Y2hhdC...rR==
AES-256-GCM • Argon2id key derivation • decryptable only with your password
 
━━━ What You See ━━━
> „Für Herr Müller mit Diabetes Typ 2 empfehle ich...“
 
→ AI Provider: sees only pseudonyms. Database: only ciphertext. Only you see plaintext.
Zero Knowledge

What Does Zero-Knowledge Storage Mean?

Nobody can read your stored data

All messages, mapping tables and conversation titles are encrypted with AES-256-GCM. The key is derived from your password (Argon2id). Without your password or your recovery key, decryption is not feasible with current technology.

Not even the operator

We store no passwords, no keys, no plaintext data. The operator has no technical ability to decrypt your stored data. This is not a matter of trust – it is cryptographic design.

Even in case of server seizure

In case of government access to the server, one finds: encrypted ciphertexts and pseudonymized conversations. No plaintext data, no mappings, no readable content.

Double protection: Pseudonymization + Encryption

Even if someone could break the encryption (which is not possible with AES-256-GCM): they would find pseudonymized texts. "PERSON_001 has HEALTH_003" – worthless without the encrypted mapping table.

The Product

Why an Intelligent Proxy?

KI-Shield sits between you and the AI provider. Your team works as usual – protection happens automatically.

ML Instead of Word Lists

Our language model detects 42 PII categories – including rare names, compound diagnoses and context-dependent data that simple word lists miss.

OpenAI-compatible API

Change one line of code – your existing applications are instantly protected. Drop-in replacement for any app that currently uses the OpenAI API.

6 AI Providers, One API

OpenAI, Anthropic, Google, Groq, Mistral, DeepSeek – all through one interface. Switch providers without changing code.

Compliance-Dashboard

Complete audit log, PII statistics, user management. Provable at any time for data protection officers and auditors.

Additionally Available

Browser ZK Mode

For users who want maximum control: In Browser ZK Mode, PII detection, pseudonymization and encryption run additionally in the browser – before data even reaches the server.

How It Works

  • PII detection in the browser (25,000+ German names, IBAN, email, health data)
  • Pseudonymization before sending
  • AES-256-GCM via WebCrypto API
  • Server receives only pseudonymized + encrypted data
  • Code inspectable in DevTools – no obfuscation

Good to Know

  • Available in web chat (not via API)
  • Requires own API key (BYOK)
  • One click in settings – no installation
  • Available for all plans (incl. Free)
For Security Teams

Technical Details

Encryption

  • AES-256-GCM (Authenticated Encryption)
  • 12-byte nonce (cryptographically random)
  • 16-Byte Authentication Tag
  • Key-Derivation: Argon2id (256-bit)
  • Format: gcm$ + Base64(nonce + ciphertext + tag)

PII Engine

  • 42 PII categories (incl. Art. 9 & 10 GDPR)
  • ML language model for context-based detection
  • Consistent pseudonym mapping per conversation
  • Re-identification only for the sender
  • No plaintext logging

API

  • OpenAI-kompatible REST-API
  • Drop-in: base_url = "https://ki-shield.eu/api/v1"
  • Streaming support (SSE)
  • Multi-provider routing
  • Rate limiting & API key management

Infrastructure

  • 8 security layers (WAF, CrowdSec, Fail2Ban, rate limiting)
  • Container mit cap_drop: ALL
  • Swap & core dumps disabled
  • Hosting in Germany (Hetzner)
  • German company, German privacy standards
FAQ

Frequently Asked Questions

It means: Everything that is stored is encrypted – and only you have the key. Messages, mapping tables, conversation titles. Neither the operator nor an attacker nor an authority can decrypt this data. The key exists only in your head (your password) or in your recovery key.

Because you would send personal data directly to servers in the USA. No audit log, no evidence, no protection. KI-Shield pseudonymizes automatically – the AI provider never sees real data, and you have a complete compliance dashboard.

Yes. KI-Shield is a Drop-in-Replacement for the OpenAI API. Change the base URL to ki-shield.eu/api/v1 – done. Works with Python, JavaScript, Go, any language with an HTTP client. Streaming is supported.

An additional option in the web chat. PII detection, pseudonymization and encryption run additionally in your browser before data is even sent. For users who want maximum control. Requires your own API key (BYOK).

An authority would find: encrypted ciphertexts, pseudonymized conversations, hashed passwords. No plaintext data. Without the respective user password or recovery key, decryption is not possible – not even for us as the operator.

Secure AI Usage for Your Enterprise

ML-based PII detection, automatic pseudonymization, AES-256-GCM encryption. One API – all AI providers – your data protected.

No credit card 50 requests/month free OpenAI-compatible API Made in Germany