Protect your data before using ChatGPT
Complete data protection suite: DLP 2.0 with 79+ patterns (10 domains, 33 categories), contextual tokenization, indirect injection detection in documents, anti-data poisoning protection, and compliance engine for nLPD, GDPR, FINMA, HIPAA, SOC 2.
DataShield is the core DLP engine of Adlibo Guard. It is included in all AG plans (Cloud and On-Prem). It is no longer available separately.
View Adlibo Guard plansEvery prompt sent to an LLM can leak your sensitive data. DataShield protects them BEFORE they reach the LLM. Available standalone (SMBs) or as Prompt Guard addon (developers).
Microsoft Copilot's sensitivity labels were bypassed twice in 8 months (CVE-2025-32711, CW1226324). DataShield tokenization protects your data regardless of the LLM used.
3,900+
DLP Patterns
<50ms
Tokenization latency
52
LLMs supported
6
Languages
FR, EN, DE, IT, ES, PT
99.99%
Availability
For SMBs using ChatGPT, Claude, Gemini directly
Route your LLM calls through our Swiss proxy. Automatic tokenization before sending, transparent rehydration on return.
For developers with custom chatbot or application
Integrate DataShield directly into your code. Full control over when and how tokenization occurs.
tokenize() before sending to LLMrehydrate() after LLM responseFor web interface protection (Copilot, ChatGPT, Claude)
Chrome/Edge extension that automatically tokenizes your data before submission to LLM interfaces.
Integrate in just a few lines of code.
# Tokenize sensitive data before sending to LLM
curl -X POST https://api.adlibo.com/api/v1/dlp/analyze \
-H "Authorization: Bearer YOUR_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"text": "Send CHF 50000 to IBAN CH93 0076 2011 6238 5295 7",
"mode": "tokenize"
}'
# Response: { "tokenized": "Send CHF 50000 to [TOKEN:xxxx]", ... }
# Rehydrate after LLM response
curl -X POST https://api.adlibo.com/api/v1/dlp/analyze \
-H "Authorization: Bearer YOUR_API_KEY" \
-d '{ "text": "Transfer confirmed to [TOKEN:xxxx]", "mode": "rehydrate", "sessionId": "sess_abc123" }'Remote work, travel, and BYOD create blind spots for network-based DLP solutions. With 67% of employees using AI outside the corporate network (Gartner 2025), your protection must not depend on a network perimeter.
Office | Remote | BYOD | Mobile | |
|---|---|---|---|---|
| Endpoint Shield100% | ||||
| SDK Integrated | ||||
| Cloud Proxy (MDM) | ||||
| Cloud Proxy (PAC/DHCP) | ||||
| On-Premise Proxy |
Only Endpoint Shield guarantees 100% protection regardless of network
The only DLP protection that follows your employees everywhere
Each deployment mode covers different layers. Combine them for defense in depth.
| Endpoint Shield Extension | DS Cloud Proxy | DS On-Premise | |
|---|---|---|---|
| User Interface (CSS / banner) | |||
| Senseway redirect | |||
| AI input blocking (CSS) | |||
| Network (HTTP blocking) | |||
| DLP tokenization | |||
| Contextual RBAC/ABAC |
Endpoint Shield is the only mode that can inject CSS into the browser (banner, input blocking). Cloud Proxy and On-Premise block at the network level.
Automatically route your LLM calls through Adlibo for transparent protection. Curated network of 10 enterprise-grade providers, each fully tested: authentication, streaming, tokenization, rehydration, automatic failover.
PAC file for automatic system configuration, no code changes required.
Environment variables to redirect SDKs to the proxy.
Simply change the baseURL in your SDK initialization.
Circuit breaker with automatic failover to backup providers.
Layer 7 application proxy that inspects and protects LLM API JSON content.
Detailed logs of each request with timestamp, user and protected data for compliance.
Compatible with all official SDKs (OpenAI, Anthropic, Google, Mistral, etc.).
Single configuration point for all your LLMs, simplifying multi-provider management.
100% perimeter coverage: intercepts all outbound LLM traffic — programmatic API calls, CI/CD pipelines, backend services. Automatic DataShield tokenization across all major AI providers (OpenAI, Anthropic, Google, Mistral, DeepSeek, xAI, etc.). Zero code change required.
10 enterprise-grade providers:
Need a specific provider? Additional providers on request (Enterprise, 48h SLA).
Modern data protection relies on a constant trade-off between utility and confidentiality. Understanding these three approaches is essential for choosing the right strategy.
| Concept | Method | Reversibility | GDPR Status |
|---|---|---|---|
Anonymization | Irreversible removal of identifying link | Impossible | Outside GDPR |
Pseudonymization | Replacement by an alias (unique ID) | Possible (via key/table) | Personal data |
Obfuscation | Reversible transformation preserving format | Possible (via algorithm) | Personal data |
Tokenization | Random token stored in a vault | Possible (via Token Vault) | Personal data |
Anonymization is the only method that frees the company from strict legal constraints. It's essential for Open Data or sharing public statistics. However, it's complex to implement: aggregating ages by ranges reduces AI model accuracy.
Pseudonymization is the ultimate operational compromise. It preserves individual "finitude": you know that user ID-99X made three purchases, without knowing their civil identity. Ideal for cohort analysis.
Obfuscation scrambles data while preserving its structure and format (e.g., email → a*****@d*****.com). Ideal for test environments or sharing with third parties without revealing actual values.
Tokenization goes one step further in security. The token has no intrinsic value without access to the centralized secure server. It's the standard for banking transactions: even if a hacker steals tokens, they can't do anything with them.
DataShield uses tokenization to protect your data before sending to LLMs. Your IBANs, emails, and sensitive data are replaced with valueless tokens, then restored in the response. The best of both worlds: maximum security and preserved utility.
Each tokenization session generates a unique key encrypted by a master key derived via scrypt. Tokens are stored in an AES-256-GCM encrypted vault with built-in authentication (GCM), ensuring both confidentiality and integrity.
Master key (env) -> Session key (unique) -> Encrypted values. Double layer protection.
AES-256-GCM ensures both confidentiality and data integrity. Any tampering is detected.
Token mappings expire automatically (24h default). Periodic cleanup via CRON.
Every vault access is logged: user, role, department, result (granted/denied).
In compliance with GDPR (Art. 4(5)) and nLPD (Art. 5), DataShield offers both approaches. Anonymization permanently removes any identifying link (outside GDPR scope), while pseudonymization (tokenization) maintains a reversible link via the secure vault.
Recommended for: public statistics, research, Open Data
Recommended for: LLM interactions, internal analytics, BI
Tokenization of patient identifiers (NIR, AVS), ICD-10 codes, lab results with statistical preservation of numerical values. Native HDS and HIPAA compliance.
Protection of case numbers, court references, and confidential amounts. Document integrity ensured via SHA-256 hash to maintain probative force.
Protection of banking data, trading accounts, and compliance indicators. Amounts are statistically preserved for portfolio analytics. Native FINMA and AMLA compliance.
Intellectual property protection: patents (EP, US, CH, PCT), trade secrets, source code snippets, and R&D projects. Trade secrets are systematically blocked.
Sensitive numerical values (amounts, lab results, scores) are encrypted with Format-Preserving Encryption (FPE). Magnitude and distribution are maintained for analytical operations.
Salary CHF 85'347 -> CHF 85'000 (rounded, computable)Over 79 DLP patterns covering 10 domains and 33 categories. Indirect injection detection in documents and anti-data poisoning protection on LLM responses.
Machine learning algorithms trained on millions of patterns to detect sensitive data with precision.
Automatic masking of sensitive data (IBAN, card numbers, emails) before transmission to LLMs.
Define your own detection patterns for data specific to your business.
Visualize where your data goes with complete data flow mapping.
Collision-free detection: every DLP pattern is automatically validated for conflicts, token format safety and performance (ReDoS). 99%+ roundtrip integrity.
Granular access control by role, department and group. AD/LDAP integration. 30+ predefined roles and partial masking.
Indirect injection detection in documents and anti-data poisoning protection on LLM responses.
Documents (PDF, DOCX, emails) can contain malicious instructions invisible to the user but interpreted by the LLM. DataShield scans and neutralizes these injections.
LLM responses can be manipulated to return corrupted or malicious data. DataShield validates outputs before delivery.
79+
DLP Patterns
10
Business Domains
33
DLP Categories
<5ms
Scan Latency
Dynamic policies, attribute-based control (ABAC), session context, and just-in-time access for granular tokenization security.
Rules that adapt in real-time based on context: time, location, device, behavior.
Access control based on user, resource, and environment attributes.
Continuous session context evaluation to detect anomalies and adjust permissions.
Just-in-time access for sensitive operations with automatic expiration.
Framework templates, gap analysis, automated evidence collection, and risk scoring to facilitate your audits.
Pre-configured templates for nLPD, GDPR, FINMA, HIPAA, and SOC 2 with control mappings.
Automatic gap analysis between your current posture and regulatory requirements.
Automated evidence collection for your audits with timestamps and guaranteed integrity.
Real-time risk score with trends and improvement recommendations.
Accelerate repetitive prompts with intelligent caching
Enterprise chatbots and standardized requests often generate similar prompts. Prompt Caching detects these repetitions and instantly returns cached results, reducing latency and API costs.
Instant response for already tokenized prompts. Perfect for FAQ, customer support and standardized queries.
Reduce your API costs by avoiding retokenization of identical prompts.
Secure hashing algorithm to quickly identify similar prompts with LRU policy.
Adjustable cache duration from 15 minutes to 24 hours based on your business needs.
Behavior-based adaptive protection
Intelligent protection that adapts to each user's behavior. Less friction for trusted users, enhanced protection for risky behaviors.
Confidence score 0-100 calculated from behavior: frequency, PII, hours, volume. Recalculated daily.
4 tokenization levels (minimal, standard, strict, full) automatically adjusted based on user profile.
Configurable learning window (7, 14, 30 days). The system improves with every interaction.
Users can report false positives to refine their profile and reduce friction.
Available with DataShield Business and Enterprise
DataShield is no longer available separately. It is an integral part of the Adlibo Guard platform. Choose your AG edition to benefit from DataShield, OCR, Shadow AI Radar and more.
Comparison with existing solutions
| Feature | DataShield | Lakera | Arthur AI | Guardrails AI |
|---|---|---|---|---|
| Swiss Hosting | - | - | - | |
| Native nLPD compliance | - | - | - | |
| DLP + contextual access control | - | - | ||
| Built-in compliance engine | - | - | - | |
| Beyond CLOUD Act | - | - | - | |
| Cloud Proxy (10 LLMs) | - | - | - |
DataShield reduces your compliance costs and risks
Reduction in audit preparation time
Average cost of avoided breach (IBM 2024)
Maximum GDPR fine avoided
Security team productivity gain
Select a context (role, industry, framework) and watch how DataShield tokenizes sensitive data before sending to LLM, then rehydrates it in the response.
Compare side by side: your exposed data on the left, your data protected by tokenization on the right.
Drag a file or use the selected scenario (.txt, .pdf, .docx, .doc, .rtf, .csv, .json, .md, .html, .xml, .xlsx, .xls, .pptx, .ppt, .odt, .ods, .odp, .jpg, .jpeg, .png, .gif, .webp, .bmp, .svg, .heic, .heif, .avif)
The LLM sees all your sensitive data
Chargement des données depuis la base SwissBank SA......
Without Protection
The LLM sees all your sensitive data (IBAN, names, emails, etc.). Risk of leaks through learning or logs.
Protected by DataShield
The LLM never sees your real data. Contextual tokenization and automatic rehydration.
Get your free API key in 2 minutes. 1,000 requests/month included, no credit card required.
Get my free API keyNeed an air-gapped deployment? DataShield On-Premise embeds complete tokenization, AES-256-GCM vault and vertical RBAC in a read-only Docker container. Your data never leaves your infrastructure.
Access 50+ AI platforms in an end-to-end protected environment. Prompt Guard detects injections, DataShield tokenizes your sensitive data, and the intelligent router selects the best model for your need.
Everything you need to know about DataShield in production.
Cloud Proxy: for SMEs using ChatGPT/Claude/Gemini directly (zero code, PAC file or env var). SDK: for developers with custom chatbot or app (full control via tokenize()/rehydrate()). Extension: for browser-level workstation DLP protection. All three can be combined for defense in depth.
Tokens are stored in the AES-256-GCM encrypted Token Vault. Retention is configurable: 24h default, extendable up to 90 days for audit needs. In On-Premise mode, the vault is on your infrastructure and you fully control retention policy.
DataShield covers 350+ DLP patterns across 14 domains and 85 categories: Finance (IBAN, credit cards), PII (names, emails, phones), Healthcare (patient numbers, diagnoses), Tech (passwords, API keys, tokens), HR (salaries, reviews), Legal (contracts, attorney privilege), Government (SSN, passports), Retail, Manufacturing, Education, Real Estate, Crypto, Insurance, Communication. Patterns customizable per organization.
Yes. DataShield works with 52 LLMs supported by the ADLIBO platform: OpenAI, Anthropic, Google, Mistral, Meta, DeepSeek, Cohere, and all models accessible via OpenRouter. Cloud Proxy automatically handles routing and failover between providers.
After the LLM generates its response containing tokens ([xxxx]), DataShield replaces them with original values by querying the Token Vault. The process is bidirectional: tokenization before sending, rehydration on return. Roundtrip similarity is 99%+.
Yes, at Business and Enterprise tiers. You can: enable/disable pattern categories, add custom patterns (regex), define RBAC role exceptions, configure sensitivity thresholds per department, and create industry/country-specific policies.
Yes. Tokenization latency is under 50ms. SHA-256 cache accelerates repeated prompts (5-10x). The complete pipeline (PII detection + tokenization + LLM send + rehydration) adds less than 100ms to total response time.
Classical encryption makes data unreadable. Tokenization replaces data with random values preserving context for the LLM. Example: "John Smith lives in Geneva" becomes "[TOKEN:042] lives in [TOKEN:017]". The LLM understands the structure but never sees real data.
DataShield is now included in Adlibo Guard. Visit the Adlibo Guard page for current pricing. For 50 employees with moderate use (200 req/day), the AG Pro plan covers needs. 14-day free trial, no credit card.
Yes. When a prompt contains search intent ("give me information about...", "search...", "go to..."), DataShield automatically disables tokenization to avoid tokenizing brand names or legitimate search terms. Intent detection is intelligent and configurable.
Yes. DataShield On-Premise embeds complete tokenization, AES-256-GCM vault and vertical RBAC in a read-only Docker container. Deployable air-gapped, Kubernetes compatible. Latency < 5ms. Ideal for regulated environments (FINMA, healthcare, defense).
Two methods: (1) PAC file — configure proxy in browsers via GPO. (2) Environment variables — HTTPS_PROXY in your applications. No code changes needed. The proxy intercepts LLM API calls, tokenizes, relays, and rehydrates the response.
Organizations hesitate to adopt AI due to data leakage fears and regulatory risks. DataShield removes each of these barriers by tokenizing data BEFORE it reaches the LLM.
DataShield tokenizes your sensitive data before sending to the LLM, then restores values in the response. Your data never leaves your perimeter.
nFADP, GDPR, FINMA, EU AI Act: compliance is built-in by design. 100% Swiss hosting, beyond CLOUD Act.
Every tokenization is logged: mappings, PII categories, timestamps. Centralized dashboard for compliance teams.
Adaptive Shield adjusts tokenization level based on user trust score (0-100). Higher risk means stricter protection.
Contextual tokenization preserves text meaning for the LLM. Rehydrated responses are 99%+ faithful — zero impact on quality.
Cloud Proxy for SMBs, SDK for developers, browser extension for everyone. 3 modes, one protection.
DataShield is no longer available separately. It is an integral part of the Adlibo Guard platform. Choose your AG edition to benefit from DataShield, OCR, Shadow AI Radar and more.
DataShield is built on industry standards to ensure interoperability, auditability and regulatory compliance.
Adhering to standards ensures interoperability with your existing tools and simplifies compliance audits.