GateWall Developer & Security Docs
Comprehensive guide to integrate GateWall as a drop-in replacement, configure in-flight DLP redaction, and ensure enterprise compliance.
1. Quickstart & Drop-in Integration
Switch from official OpenAI API to GateWall in 1 line of code. Zero refactoring required.
curl https://gatewall.eu/v1/chat/completions \ -H "Content-Type: application/json" \ -H "Authorization: Bearer gw_live_YOUR_VIRTUAL_KEY" \ -d '{ "model": "mistral-large-4", "messages": [ { "role": "user", "content": "Audit client Alice Martin, email: alice@banque.fr, IBAN: FR7630006000011234567890189." } ] }'GateWall exposes the exact OpenAI API schema (/v1/chat/completions). You do not need to install custom SDKs or rewrite prompts.
2. In-Flight DLP & PII Redaction Engine
Automatic regex and NLP masking executed in volatile RAM before prompt transmission to upstream providers.
Concrete In-Flight Encoding & Decoding Walkthrough
See how GateWall intercepts raw prompts, masks sensitive data into tokens before sending to the LLM, and restores original data on egress in volatile RAM.
Anatomy of an In-Flight Token
GateWall uses syntactically isolated tokens that frontier LLMs naturally treat as typed variables without hallucinating.
Mathematical angle brackets prevent token fragmentation in standard BPE tokenizers.
Preserves grammatical role (NAME, EMAIL, IBAN) so the model maintains accurate syntax.
Maintains referential consistency when the same entity is referenced across multiple sentences.
Cryptographic 16-bit salt generated per stream in RAM, preventing cross-session correlation.
Volatile Memory Pipeline (< 20ms overhead)
Zero Disk Write · In-Memory MapClient sends raw prompt with sensitive customer data to GateWall.
Identifies PII (email, IBAN, SSN, IP, NDA) and swaps with reversible tokens in RAM.
OpenAI / Mistral receives only anonymized tokens. Zero plaintext exposure.
Upstream response is rehydrated with original data before returning to client.
Supported PII Entites & Tokens
7 Built-in Detectors| Data Type | Sample Input | Redacted In-Flight Token | Detection Method | Reversible on Egress |
|---|---|---|---|---|
| Email Addresses | alice.dupont@bank.com | ⟪AG:EMAIL:1:5ab0⟫ | RFC 5322 Regex + Domain Check | Yes (RAM-only) |
| Phone Numbers (FR & Int.) | +33 6 12 34 56 78 | ⟪AG:PHONE:1:5ab0⟫ | E.164 + National Formats (FR/EU/US) | Yes (RAM-only) |
| Bank Account (IBAN / RIB) | FR7630006000011234567890189 | ⟪AG:IBAN:1:5ab0⟫ | ISO 13616 Modulo-97 Verification | Yes (RAM-only) |
| Credit Cards (PAN & CVV) | 4532 •••• •••• 8891 | ⟪AG:CARD:1:5ab0⟫ | Luhn Algorithm (Visa, MC, CB, Amex) | Yes (RAM-only) |
| Social Security (NIR / SSN) | 1 85 12 75 108 245 42 | ⟪AG:SSN:1:5ab0⟫ | INSEE NIR 15-digits & US SSN | Yes (RAM-only) |
| Internal IP Addresses | 192.168.1.42 / 10.0.4.1 | ⟪AG:IP:1:5ab0⟫ | IPv4 Private & Public CIDR | Yes (RAM-only) |
| Custom NDA Project Names | Project Titan / QuantumVault | ⟪AG:PROJECT:1:5ab0⟫ | Exact & Soundex Match | Yes (RAM-only) |
3. Supported by Open Source Core
Zero-dependency, MIT-licensed in-flight redaction engine running in volatile RAM.
Powered by avantGate
GateWall's zero-leak in-flight redaction and reversible tokenization algorithms run on avantGate, our auditable open-source engine. Inspect the code, audit detection patterns, or use the standalone library in your local pipelines.
Install via npm and anonymize prompts in 3 lines of code directly inside your local application runtime:
import { createInFlightRedactionSession } from "avantgate";// 1. Initialize an ephemeral in-memory session (RAM only)const session = createInFlightRedactionSession();// 2. Anonymize user prompt before calling any upstream LLMconst rawPrompt = "Audit user Alice Martin (alice@corp.eu), IBAN: FR7630006000011234567890189";const maskedPrompt = session.mask(rawPrompt);console.log(maskedPrompt);// Output: "Audit user Alice Martin (⟪AG:EMAIL:1⟫), IBAN: ⟪AG:IBAN:1⟫"// 3. Upstream model processes only synthetic tokens...// Then restore plaintext on egress response:const restoredText = session.restore("Confirmed transfer for IBAN: ⟪AG:IBAN:1⟫");console.log(restoredText);Choose the right architecture based on whether you need a local in-process library or a managed team gateway.
| Capability / Feature | avantGate (OSS Library) | GateWall (Managed Gateway) |
|---|---|---|
| Deployment Model | In-process SDK (npm i avantgate) | Universal HTTP Proxy (Zero SDK, drop-in OpenAI) |
| In-Flight PII Redaction | Local RAM buffers (in-process) | Volatile RAM (< 20ms, purged per stream) |
| Multi-Provider Routing | Self-implemented | Universal (OpenAI, Claude, Mistral) + HA Fallback |
| Virtual Keys & Budget Quotas | None | Isolated Virtual Keys + HTTP 402 Circuit Breaker |
| FinOps Telemetry & Observability | None | Realtime Cockpit, p95 latency, token spend & margins |
| Cryptographic DPO Audit | None | HMAC-SHA256 Audit Proofs & GDPR Art. 28 DPA |
| Best Suited For | Solo devs, local pipelines, R&D scripts | B2B teams, SMEs, scale-ups, DPO & production |
4. Zero-Data Retention & DPO Cryptographic Proof
Proof of zero disk storage and tamper-evident audit certificates for GDPR and AI Act compliance.
Prompt and completion payloads strictly exist in volatile memory buffers for the duration of the stream. Once the HTTP stream terminates, RAM buffers are immediately flushed. No disk writes, no vector databases, no model training.
Every session receives a cryptographically verifiable HMAC-SHA256 signature containing token count, masked PII types, and latency metrics for compliance audits without storing payload content.
5. Trust Center, Legal & DPO Compliance Guide
Clear legal guarantees, processor duties under GDPR Art. 28, zero-training commitments, and real encryption architecture.
GateWall acts strictly as a data processor on behalf of your organization. You remain the sole Data Controller and retain exclusive ownership over your business data. A standardized Data Processing Agreement (DPA) compliant with EU SCCs is readily provided.
Customer prompts and completions are never sold, shared, or used to fine-tune foundation models. In-flight tokenization guarantees that no identifying records leave your secure EU boundary.
All client traffic is strictly enforced over TLS 1.3 with Perfect Forward Secrecy. Upstream BYOK keys are encrypted at rest with hardware-accelerated AES-256-GCM. Session audit proofs are cryptographically signed via HMAC-SHA256.
Sensitive data (SSN, payment cards, phone numbers, IP addresses) are neutralized into synthetic tokens before reaching external models. For maximum confidentiality, the 'Zero-Trace' mode guarantees instant volatile RAM destruction with zero payload persistence.
Review our standard terms or contact our legal team at legal@gatewall.eu for our EU Standard Contractual Clauses (SCC) Data Processing Agreement.
6. Agent Guardrails (Tool Health & HITL)
Prevent runaway agent loops and supervise high-risk autonomous actions in production.
Anti-Loop Circuit Breaker (OWASP LLM06)
Monitors tool execution sequences. When an AI agent invokes the same tool with identical or oscillating parameters more than 3 times, GateWall trips an immediate circuit breaker to protect budgets.
Human-in-the-Loop Supervision (OWASP LLM03)
Critical actions (database migrations, fund transfers, email blasts) are paused in PENDING state awaiting human operator validation via the Cockpit or API webhook.
7. Universal API Reference
OpenAI-compatible endpoints served by GateWall.
Universal chat completions proxy with in-flight DLP, spend controls, and automatic fallback.
Catalog of 100+ available LLM models with pricing and context windows.