ZENTRIC PROTOCOL

PII Integrity · Deterministic Infrastructure · Secure Protocol

The protocol layer between intent and execution in AI systems.
Every signal examined. Every verdict signed. Nothing passes without record.

→ Request Access · Documentation · Integrity Report v1.0

Repository Scope & Commercial License

This repository exists for transparency and contribution — not as a deployable alternative to the hosted service.

Cloning this repository does not give you access to the Zentric processing service. The detection engine that inspects prompts, detects PII, and generates signed audit reports runs on Zentric's infrastructure and requires an active license.

Why publish the middleware?

Because trust is infrastructure. You should be able to verify how authentication works, how your API key is validated, and how subscription state is checked before your requests reach the engine. We believe in auditability at every layer — including our own enforcement code.

Contributions welcome

We accept contributions to the middleware, webhook handler, and Supabase schema. Open a PR or file an issue. For security-related contributions, see the Security section.

Getting access

What is Zentric Protocol?

Zentric Protocol is an infrastructure integrity layer for AI systems. It sits between your application and your LLM, examining every signal — prompts, responses, user inputs — and returning a cryptographically-signed verdict before execution continues.

It is not a filter. It does not guess. It applies deterministic rules across a standardized pipeline and returns a structured, auditable JSON report for every request.

Input Signal
     │
     ▼
┌─────────────────────────────────────────┐
│           ZENTRIC PROTOCOL              │
│                                         │
│  ┌─────────────┐  ┌─────────────────┐  │
│  │IntegrityGuard│→│  PrivacyGuard   │  │
│  │ 22 injection │  │  17 PII types   │  │
│  │  signatures  │  │  7 languages    │  │
│  └─────────────┘  └────────┬────────┘  │
│                             ▼           │
│                    ┌──────────────┐     │
│                    │ ZentricReport│     │
│                    │ UUID+SHA-256 │     │
│                    │  GDPR Art.30 │     │
│                    └──────────────┘     │
└─────────────────────────────────────────┘
     │
     ▼
Verdict + Certificate → Your System

Performance Benchmark

Extracted from Zentric Integrity Report v1.0 — 1,000,000 simulations across all supported attack vectors and entity types.

Full methodology and raw data available on request: core@zentricprotocol.com

Benchmark Methodology

The 1,000,000-simulation corpus was constructed from four sources: (1) published prompt injection research datasets (PINT Benchmark, PromptBench, garak); (2) adversarial samples hand-authored across 22 attack categories to stress-test edge cases not present in public datasets; (3) synthetically mutated variants of known attack patterns — character substitution, whitespace injection, Unicode normalization attacks, and mixed-language payloads — to measure robustness against obfuscation; and (4) benign control samples drawn from production-representative traffic to verify precision does not degrade under normal use. The final split is approximately 53% attack samples and 47% benign controls. Simulations were run deterministically: the same corpus against a frozen signature set, with no retraining or tuning between runs.

What this benchmark does not cover: adversarial inputs specifically constructed to defeat these 22 signatures after reading this repository (signatures are partially public, which is a known trade-off of transparency); semantic prompt injections that do not match any current signature pattern and rely entirely on model misinterpretation; non-English languages beyond the seven supported (EN, ES, FR, DE, IT, PT, NL); and multi-turn conversation-level attacks where the injection is distributed across several messages. The precision figures above reflect deterministic pattern matching — not an ML classifier, not a generalization claim. Zentric is honest about what it detects and what it does not.

The Three Modules

01 · IntegrityGuard

Detects prompt injection, jailbreak attempts, and instruction overrides before they reach your LLM.

• 22 catalogued injection signatures

• 7 supported languages (EN, ES, FR, DE, IT, PT, NL)

• Multilingual NLP classification layer

• Mean server-side processing: 23.4ms (network latency not included)

02 · PrivacyGuard

Identifies and anonymizes PII in prompts and responses. Regional standards treated as first-class entities.

• 17 PII entity types: SSN, NIF, CPF, CURP, IBAN, SWIFT, passport, email, phone, and more

• Regional pattern recognition (EU, US, LATAM)

• Anonymization operators: redact, mask, tokenize, pseudonymize

• Recall rate: 99.71% across 17 entity types

03 · ZentricReport

Every request that passes through the protocol generates a signed, immutable audit record.

{
  "report_id": "zp_01HXYZ...",
  "uuid": "f47ac10b-58cc-4372-a567-0e02b2c3d479",
  "timestamp_utc": "2026-05-14T22:00:00.000Z",
  "sha256": "e3b0c44298fc1c149afb...",
  "verdict": "CLEARED",
  "integrity": {
    "injection_detected": false,
    "signatures_matched": [],
    "confidence": 0.9998
  },
  "privacy": {
    "pii_detected": true,
    "entities": [
      { "type": "EMAIL", "action": "REDACTED", "position": [42, 61] }
    ]
  },
  "compliance": {
    "gdpr_art30": true,
    "ccpa": true,
    "eu_ai_act_s52": true
  },
  "latency_ms": 21.4
}

API Reference

Authentication

curl -X POST https://api.zentricprotocol.com/v1/analyze \
  -H "Authorization: Bearer zp_live_..." \
  -H "Content-Type: application/json" \
  -d '{
    "input": "Your prompt or user input here",
    "modules": ["integrity", "privacy"],
    "options": {
      "anonymize": true,
      "language": "auto"
    }
  }'

Response

{
  "status": "ok",
  "verdict": "CLEARED",
  "report": { ... },
  "anonymized_input": "Your prompt or user input here",
  "latency_ms": 23.1
}

Verdict States

SDKs

Compliance Coverage

Zentric Protocol is designed from the ground up for regulated AI deployments.

Pricing

→ See plans · → Get API key · → Contact for Enterprise

Architecture Principles

Deterministic. The same input always produces the same verdict. No probabilistic black boxes in the critical path.

Stateless. The protocol does not store your data. Each request is processed and returned. The audit record is yours.

Composable. Deploy the full stack, a single guard, or wire only the audit layer into existing infrastructure.

Auditable. Every verdict is signed with SHA-256, timestamped in UTC, and assigned a UUID. Your compliance team will thank you.

For Agent Pipelines

Agent attacks don't arrive through the chat input. They arrive through tool call responses, RAG chunks, and memory retrievals — any external content that enters the prompt window. Your system prompt doesn't protect you here: it doesn't run until after the input is already parsed.

Wire Zentric at every ingestion point, not just on user messages:

• LLM input — user messages before they reach the model

• Tool output — external API responses before they re-enter the context window

• RAG retrieval — document chunks before they are assembled into the prompt

• Memory reads — stored context before it is injected into the next turn

One POST to /v1/analyze. The verdict comes back in ~23ms. The agent continues or halts based on the result. Nothing else changes in your pipeline.

curl -X POST https://api.zentricprotocol.com/v1/analyze \
  -H "Authorization: Bearer zp_live_..." \
  -H "Content-Type: application/json" \
  -d '{"input": "<tool_output_or_rag_chunk_here>", "modules": ["integrity", "privacy"]}'

MCP Server — Claude Desktop Integration

Zentric Protocol ships a native Model Context Protocol (MCP) server that integrates directly with Claude Desktop and any MCP-compatible agent runtime.

What it does

The MCP server exposes Zentric's detection engine as a native MCP tool. When wired into Claude Desktop, the agent automatically calls analyze_prompt before sending any input to the LLM — user messages, tool responses, RAG chunks, and memory retrievals are all checked.

MCP Tool exposed

analyze_prompt(text: string) -> ZentricReport

Returns: verdict (CLEARED / BLOCKED), risk_score, matched signatures, pii_entities, report_hash (SHA-256), latency_ms.

Install via npm

npx zentric-protocol-mcp

Claude Desktop configuration

Add to your claude_desktop_config.json:

{
  "mcpServers": {
    "zentric-protocol": {
      "command": "npx",
      "args": ["zentric-protocol-mcp"],
      "env": {
        "ZENTRIC_API_KEY": "your_api_key"
      }
    }
  }
}

Get your API key at zentricprotocol.com/quickstart — free tier is 10,000 requests/month, no credit card required.

MCP server source

The MCP server source code is in /mcp-server. It is built with the Model Context Protocol SDK and published to npm as zentric-protocol-mcp.

Security

We take the security of this protocol seriously. If you discover a vulnerability, please report it responsibly.

• Email: core@zentricprotocol.com

• Subject: [SECURITY] <brief description>

• Response SLA: 48 hours acknowledgement, 7 days resolution target

We do not operate a public bug bounty program at this time. Responsible disclosure is acknowledged in our changelog.

Roadmap

• IntegrityGuard v1.0 — 22 signatures, 7 languages

• PrivacyGuard v1.0 — 17 PII types, EU/US/LATAM

• ZentricReport v1.0 — SHA-256, UUID, GDPR Art.30

• REST API (production)

• Python SDK — Q3 2026

• Node.js SDK — Q3 2026

• Streaming support (SSE) — Q3 2026

• Webhook callbacks — Q4 2026

• SOC 2 Type II certification — Q4 2026

• Self-hosted deployment option — 2027

Contact

Zentric Protocol · Infrastructure Integrity for the AI Era

zentricprotocol.com · © ZP MMXXVI · v1.0.0

Built for CTOs who know that trust is infrastructure, not a feature.