Anonymize Data Before It Reaches the LLM
Every AI prompt, every RAG pipeline query, every LLM fine-tuning dataset — if it contains PII, it's a compliance risk. We remove it before it leaves your control.
The Risk: PII Inside AI Workflows
AI data safety is the practice of removing personally identifiable information (PII) from prompts, RAG queries, and training data before a large language model can process it. Large language models ingest everything you send them. Once PII enters an LLM, you lose control over how it is stored, processed, or surfaced.
- Training Data Contamination — Personal data may be absorbed into future model weights
- Prompt Logging — AI providers may log inputs for debugging, safety review, or model improvement
- Output Leakage — PII from one user's prompt can appear in another user's response
- No Delete Capability — Once data enters a model, there is no reliable way to remove it
- Cross-Border Transfer — AI API calls may route data through non-EU jurisdictions
Pre-Processing Pipeline
A deterministic pre-processing layer that reduces PII exposure before data reaches an AI service — with reversible tokens for re-identification when authorized.
The AI never sees real names, emails, or sensitive data — it works with safe tokens that preserve context and meaning.
MCP Server — AI Safety Built Into Your IDE
The Model Context Protocol (MCP) integrates anonymize.solutions directly into Claude Desktop, Cursor, and VS Code. The MCP server provides anonymization tools that the assistant calls when needed — it is not an automatic filter in front of every input, so text you type or paste directly into the chat reaches the model as written.
Setup in 5 Minutes
Add the MCP Server to your AI tool's configuration file. One JSON block, one API key — protection starts immediately for every conversation.
Claude, Cursor, VS Code
Works with Claude Desktop, Claude Code, Cursor IDE, Windsurf, and any MCP-compatible AI assistant via Streamable HTTP transport.
Processing, and what Zero-Knowledge covers
Your text passes through, gets anonymized, and returns — the detection service necessarily reads the text it analyses. Where it may not be read at all, the local client keeps it on your own machine.
The assistant calls an MCP tool on your text → PII replaced with safe tokens → LLM processes clean data → response de-anonymized automatically.
Three ways to protect AI workflows
Choose the integration point that matches your workflow — or combine all three for end-to-end coverage.
MCP Server
For developers and technical teams. Integrates directly into your IDE and gives the assistant tools to anonymize text before working with it.
- Claude Desktop, Cursor, VS Code
- Anonymization on the assistant's tool call
- 9 MCP tools
- Controlled data release
Chrome Extension
For business users and non-technical teams. Protects AI chat inputs on ChatGPT, Claude, and Gemini with real-time PII interception and automatic response de-anonymization.
- ChatGPT, Claude, Gemini, Copilot
- Real-time input interception
- Automatic response restoration
- Seamless browsing experience
REST API
For pipelines and automated workflows. Programmatic integration into RAG ingestion, ETL processes, ML training pipelines, and any system that sends data to LLMs.
- RAG, ETL, ML pipeline integration
- Batch processing capabilities
- JSON request/response format
- JWT authentication + rate limiting
Where AI safety matters most
Every AI workflow that processes personal data needs a pre-processing layer. These are the most common scenarios.
AI Chatbot Protection
Customer service bots processing personal data — names, account numbers, addresses, health information. Anonymize inputs before the LLM generates a response, then re-identify tokens in the output for the agent.
LLM sees: "I'm [NAME_1], my account is [ACCOUNT_1]"
RAG Pipeline Safety
Anonymize documents before vector embedding so your knowledge base holds far less PII. When the retrieval-augmented generation system surfaces context, detected personal data stays out of LLM prompts and responses — risk reduction, not a guarantee.
Query → Retrieve clean chunks → LLM
AI Agent Workflows
Multi-step agent chains where data passes through multiple LLM calls, tool invocations, and external APIs. Each step is a potential exposure point. Anonymize at the entry gate, restore at the exit.
→ Step 3 → De-anonymize → Output
LLM Fine-Tuning Prep
Clean training datasets of PII before fine-tuning. This reduces the personal data your custom model can memorize from training examples and supports GDPR data minimization and purpose limitation (Art. 5 GDPR).
→ Fine-tune model on the cleaned data
Why deterministic detection matters for AI
We use NLP + Pattern engines — not LLMs — to detect PII. This is a deliberate architectural choice with concrete benefits for AI safety.
The Problem with LLM-Based Detection
Some anonymization tools use AI models to detect PII. This creates a circular dependency: you send sensitive data to one LLM to protect it from another LLM. Our approach eliminates this paradox entirely.
Benefits for AI Workflows
- No Hallucination Risk — Pattern and NLP engines cannot invent false positives from imagination
- Complete Audit Trail — Every detection has a confidence score, entity type, and position — fully traceable
- No Data Sent to Third-Party AI — Detection happens on our EU infrastructure, never via external AI APIs
- No LLM Calls — Detection uses pattern matching and NLP models (spaCy, Stanza, transformers), not large language models
Deterministic vs. AI-Based Detection
| Property | Our Approach | LLM-Based |
|---|---|---|
| Hallucination Risk | ✓ Zero | ✗ Present |
| Audit Trail | ✓ Full | ✗ Limited |
| Data Exposure | ✓ None | ✗ To AI Provider |
| GDPR Compliance | ✓ Built-in | ✗ Depends |
KEY DIFFERENTIATORS
- Engines: Microsoft Presidio NLP + Regex Pattern
- Entities: 267 types across 48 languages
- Methods: Replace, Redact, Mask, Hash, Encrypt
- Hosting: anonym.legal on Hetzner servers in Germany; anonym.plus on your own machine
Need a comprehensive AI governance strategy?
anonymize.solutions handles the technical layer — PII detection and anonymization. For enterprise AI governance consulting, policy frameworks, risk assessments, and EU AI Act compliance, curta.solutions, the team that built this platform, provides dedicated advisory services.
GOVERNANCE
Policy + Technology
Does this apply to your organisation?
The guide is general; your obligation is not. Tell us the sector and where the data is allowed to go, and we will tell you what applies.