ZTDS Architecture Specification · 100% Local CPU Execution

Technical Security & Architecture

PrivacyScrubber enforces data confidentiality not through legal policy or backend promises, but through mathematical impossibility and local execution boundaries. All string scanning, deterministic ZTDS™ detection pattern matching, and cryptographic tokenization occur strictly inside endpoint RAM.

< 1ms
Processing Latency
Zero network roundtrips
0 Bytes
Server Ingestion
Zero backend API endpoints
Volatile RAM
Ephemeral Memory
0 disk footprint / 0 database
30 Profiles
Sanitization Rules
ZTDS™ Level-1 Certified Pioneer ImplementationZTDS-CERT-2026-PRIVACYSCRUBBER-WEB-D02DC6

Cryptographically signed Ed25519 conformance seal • IETF Internet-Draft draft-sibiryakov-ztds-protocol-01

Verify Live Certificate

Zero-Trust Data Sanitization (ZTDS) Lifecycle

How data flows exclusively within the local host security perimeter.

Zero-Trust Data Sanitization Architecture Lifecycle - 100% Client-Side Processing
Figure 1: The ZTDS execution model guarantees that plaintext PII is replaced with non-invertible surrogate tokens before dispatch to any cloud AI provider.
1. Ingest & Disambiguate

Raw inputs are evaluated locally in memory through our Contextual Entity Disambiguation Engine across 30 specialized industry compliance profiles.

30 ProfilesLocal CPU / RAM
2. Deterministic Mapping

Surrogate tokens (e.g. [NAME_1]) are deterministically assigned while preserving strict code, JSON, and table formatting integrity.

[TYPE_N] SurrogatesSyntax Preserved
3. Ephemeral Rehydration

When the AI responds with sanitized tokens, 1-Click Reveal restores original values in sub-millisecond RAM execution, with zero disk trace.

< 1ms Reveal0 Disk Trace

Zero-Database Credential Architecture — Why No Trial or Login Exists

Because PrivacyScrubber has zero central backend servers, there is physically no user database, no credential store, and no authentication API. We do not require users to create accounts, set passwords, or provide corporate emails. There is no expiring trial mechanism — the Community Tier is 100% free forever for all users directly in local browser RAM without gatekeeping. For enterprise evaluations requiring 30 specialized industry profiles, batch processing, or team governance, corporate pilots operate on strict 7-day RS256 JWT cryptographically signed tokens with automated fail-closed expiration.

Auth Model: 0 Central DB · Offline Asymmetric VerificationVerify 0 Subprocessors →

Execution Isolation Across All 4 Surfaces

Detailed technical runtime boundaries for every deployment modality.

Web Workspace (In-Browser SPA)

Web Worker Sandbox

Executes via an isolated HTML5 Web Worker thread detached from the main UI thread. Employs the ZTDS™ offline OCR and local PDF engine for local OCR. Network activity during scrubbing is physically blocked by Content Security Policy directives (connect-src 'none' for worker endpoints).

CSP: connect-src 'none'Launch Workspace →

Chrome Extension (Manifest V3)

Target Tab Isolated World

Injected strictly into active target tabs (ChatGPT, Claude, Gemini) within an Isolated World context. The in-page shield HUD (Alt+Shift+X) intercepts the text area DOM node locally. Zero background page syncing or remote logging is executed.

DOM: Closed Shadow RootInstall MV3 Extension →

PII MCP Server (@privacyscrubber/mcp-server)

Local Stdio / IPC Stream

Compliant with Model Context Protocol (Anthropic/Cursor). Communicates over local standard input/output (stdio) or loopback interface (127.0.0.1). Does not initialize HTTP listener ports, external sockets, or telemetry beacons.

Transport: JSON-RPC 2.0 StdioMCP Server Spec →

Developer SDK (@privacyscrubber/sdk)

Deterministic Node/WASM

Isomorphic npm package with zero runtime telemetry dependencies. Functions execute synchronously on input strings. License validation is achieved via offline asymmetric cryptographic signature checks (Ed25519) without phone-home verification.

Auth: Offline Ed25519 KeyExplore Node/WASM SDK →

RAM Lifecycle & Dual-Plane Cryptographic Primitives

Zero-persistence memory destruction and Protocol V5 authenticated session handoff.

Volatile RAM Token Dictionary Lifecycle

The mapping dictionary ([NAME_1] → "John Doe") is instantiated as a tab-isolated in-memory JavaScript Map object (with ephemeral tab-scoped backup for soft-reload resilience). It is strictly never committed to:

  • Persistent browser storage or disk files
  • Plaintext local encrypted browser storage or WebSQL
  • HTTP Cookies, Service Worker caches, or remote APIs

Upon tab closing or clicking "Reset Session", all references and tab-scoped memory allocations are immediately severed and zeroed.

Persistence: 0 Disk FootprintZeroized on Tab Close

Protocol V5 Dual-Plane Session Handoff

When TEAMS Session Handoff exports an encrypted dictionary via RFC 3986 URL #fragment (< 1,850 chars, 0 HTTP egress) or .pssession container, Protocol V5 (0x05 framing) enforces:

  • Payload Compression: In-memory binary stream compression prior to AEAD sealing.
  • Symmetric AEAD Cipher:XChaCha20-Poly1305 (256-bit key, 192-bit nonce, 128-bit MAC tag).
  • Memory-Hard KDF:Argon2id (64MB RAM, 3 iterations, 4 lanes; WebCrypto PBKDF2-SHA256 fallback).
  • Recipient Licensing: External recipients (outside counsel, clients, auditors) decrypt locally in 0.4s for free.
Framing: 0x05 · RFC 3986 #hashExplore TEAMS ($99/mo) →

Cryptographic Taxonomy: Client-Side ZTDS™ vs. Homomorphic Encryption (FHE) & Legacy PGP

Why lattice-based encrypted compute (FHE) and PGP fail on commercial Frontier Cloud LLMs — and how Dual-Plane ZTDS solves it in < 1ms.

1. The FHE Compute & Server Wall

Why Homomorphic Encryption (FHE) Cannot Power Frontier LLMs

Fully Homomorphic Encryption (FHE — e.g., lattice-based encrypted compute platforms) allows mathematical evaluation over ciphertexts, but incurs a 10,000× to 1,000,000× computational overhead (10s–120s+ per query) and requires custom homomorphic circuits on the server. Commercial frontier LLMs (OpenAI GPT-4o/o3, Anthropic Claude 3.7, Google Gemini) execute standard floating-point Transformer weights and do not accept FHE ciphertexts.

Latency: 10s–120s+Custom Server Circuits
2. The PGP Linguistic Blindness Wall

Why Legacy PGP & E2EE Break AI Reasoning

End-to-end encryption (OpenPGP, S/MIME, Signal) collapses the entire document into a single opaque ASCII-armored ciphertext block (-----BEGIN PGP MESSAGE-----). If pasted into ChatGPT or Claude, the LLM is completely blind to syntax, legal clauses, clinical timelines, and code structure, rendering AI reasoning impossible until decrypted.

Syntax: DestroyedOpaque Ciphertext
3. The ZTDS™ Dual-Plane Breakthrough

Sub-Millisecond Client RAM · Zero Server Cooperation

ZTDS™ requires zero cryptographic cooperation from OpenAI, Anthropic, or Google. By splitting the Sanitized AI Context Plane ([NAME_1], [IBAN_1]) from the Local RAM Dictionary Plane, standard uncooperative Cloud LLMs reason at 100% native token speeds while direct PII never leaves client RAM (< 1ms execution).

Latency: < 1ms RAM100% Native AI Reasoning
Architectural DimensionClient-Side ZTDS™ (PrivacyScrubber)Homomorphic Encryption (FHE)Confidential Computing (Server TEEs)Legacy PGP / E2EECloud DLP / Token Vaults
Works with Commercial Cloud LLMs (ChatGPT, Claude, Gemini)Yes (100% Native Reasoning)No (Requires Custom FHE Server Circuits)Only Self-Hosted Enclave ModelsNo (Opaque Ciphertext Blinds AI)Yes (Via Cloud Proxy MITM)
End-to-End Processing Latency< 1 ms (Local CPU RAM)10,000 ms – 120,000+ ms (10⁴×–10⁶× Penalty)150 ms – 800 ms (Remote Enclave RTT)5 ms (Manual CLI / Keyring)600 ms – 2,500 ms (Cloud API RTT)
Raw PII Network Egress0 Bytes (Stays in Device RAM)Encrypted Lattice Ciphertext EgressFull Raw PII Egress to Cloud EnclaveArmored Ciphertext EgressFull Raw PII Egress to DLP Vendor
Collaborative Team HandoffProtocol V5 #hash SafeLink (Recipients Decrypt Free in 0.4s)Requires Shared FHE Evaluation KeysServer IAM Session AccessManual Public Keyring & Re-TypingPaid Per-Seat Cloud DB Lookup
Statutory Compliance PostureGDPR Art. 32, HIPAA & SOC 2 Safe Harbor (Zero DPA)Cryptographic Research / Niche ComputeRequires Cloud Processor BAA / DPAStorage/Transit Only (No AI Use)Sub-Processor DPA & Breach Liability

The 3-Step "Airplane Mode" Verification Playbook

Independent technical audit protocol for security engineers and SOC analysts.

Step 1: Open Workspace

Navigate to PrivacyScrubber in Chrome, Firefox, or Safari.

Client Hydrated0 Auth Needed
Step 2: Disconnect Network

Toggle your device to Airplane Mode or physically disconnect your Ethernet cable.

Physical Air-GapNIC Disabled
Step 3: Scrub & Inspect

Paste confidential text. Notice immediate sanitization with DevTools Network panel recording 0 requests.

0 Packets Sent100% RAM Pass
DevTools Network Monitor (Live Diagnostic Probe) Active Egress Sniffer
POST /api/v1/scrub (Remote API)ERR_INTERNET_DISCONNECTED (0 bytes)
POST /telemetry/events (Telemetry)ERR_INTERNET_DISCONNECTED (0 bytes)
Local CPU Web Worker SandboxSTATUS 200 OK · 0.12ms Execution · 0 bytes Egress
Open Full DLP Latency Benchmark →

Automated QA Suite & Empirical Security Benchmarks

Continuous Integration (CI) test assertions executed on every software build.

5.68 MB
JS Memory Heap
Budget: < 35 MB (Passed)
70.8 ms
DOM Mutation Scan
5,000 Live Nodes (< 100ms)
236 ms
First Contentful Paint
Budget: < 1200ms (Passed)
0 Bytes
Network Egress
Zero prompt exfiltration
Scroll test matrix horizontally
Automated CI Test SuiteTarget Threat VectorVerified Security AssertionCI Status
E2E UI & XSS Guard DOM-based XSS & Reverse Scrub Payload Injection HTML tags sanitized to textContent; zero JavaScript execution. 100% Passed
Extension Shadow DOM Isolation Host-Page DOM Traversal & Script Hijacking Closed Shadow Root (mode: "closed"); host scripts cannot access extension UI. 100% Passed
Tab Isolation & Memory Lifetime Cross-Tab Data Contamination & Storage Persistence 0 token leakage between parallel tabs; 100% memory zeroed on reload. 100% Passed
Extension Weak Vectors Guard Prototype Pollution & Cross-Frame Message Spoofing Rejects e.source !== window; postMessage prototype pollution mitigated. 100% Passed
Cross-Platform Crypto Handoff Cryptographic Desynchronization 100% AES-GCM / Argon2id bit-level interop across Web, Extension & MCP. 100% Passed
MCP Stdio JSON-RPC 2.0 Protocol Network Port Exposure in AI IDEs (Cursor/Claude) Air-gapped stdio transport only; zero listening network ports opened. 100% Passed

Independent CLI & Socket Verification

Reproduce zero-egress cryptographic proof on your own air-gapped infrastructure.

Socket-Level Invariant AuditDeterministic

Monkey-patches node:net, node:tls, and node:http. Simulates 30 specialized profiles across 12,000+ characters to mathematically assert 0 socket openings and 0 bytes egress.

npm run verify:zero-egress
Hooks: node:net / tls / http0 Sockets Asserted
In-Process SDK Demo VerificationZero Install

Run the zero-trust sanitization engine directly in any terminal via npx. Verifies local string tokenization and sub-millisecond execution latency.

npx @privacyscrubber/sdk demo
Runtime: Node.js / WASMSDK Docs →

Academic Attestations & Protocol Standards

Formal mathematical publications, preprint archives, and IETF protocol drafts.

Architectural Verification Complete

Deploy Zero-Trust Sanitization in Your Stack

Integrate in-process Node.js/WASM sanitization into your backend ETL and RAG pipelines, or deploy company-wide via Chrome Extension and Teams workspace governance.

Technical Security & Cryptographic FAQs

Direct architectural answers for CISOs, cryptographers, and DPOs.

How does Client-Side ZTDS™ compare to Fully Homomorphic Encryption (FHE) for generative AI?
Fully Homomorphic Encryption (FHE) allows computation over lattice-encrypted ciphertexts, but incurs a 10,000× to 1,000,000× compute overhead (10s–120s+ per query) and requires custom homomorphic evaluation circuits on the server. Commercial frontier LLMs (OpenAI GPT-4o/o3, Anthropic Claude 3.7, Google Gemini) execute standard floating-point Transformer weights and do not accept FHE ciphertexts. Client-Side ZTDS™ executes in under 1 millisecond in local device RAM with zero server cooperation required, replacing direct PII with syntax-preserving surrogates ([NAME_1]) so frontier LLMs reason at native token speeds.
Why does legacy PGP or end-to-end encryption fail for collaborative AI workflows?
Legacy PGP and E2EE collapse an entire document into an opaque ciphertext block (-----BEGIN PGP MESSAGE-----). If pasted into ChatGPT or Claude, the LLM cannot parse syntax, legal clauses, or clinical logic. ZTDS Dual-Plane Session Handoff separates the Sanitized AI Context Plane from the Key Dictionary Plane (sealed via Protocol V5 Argon2id + XChaCha20-Poly1305 inside an RFC 3986 URL #fragment), allowing full AI reasoning and 0.4-second free recipient restoration.
How does client-side ZTDS™ differ from server-side Confidential Computing (TEEs / Secure Enclaves)?
Server-side Trusted Execution Environments (TEEs such as Intel SGX, TDX, or NVIDIA H100 Confidential Computing) still require transmitting raw cleartext PII over the network to a remote cloud enclave, triggering Data Processing Agreements (DPAs), cross-border transfer assessments, and remote hardware attestation dependencies. ZTDS™ sanitizes data inside local endpoint RAM before any network socket opens (0 bytes raw PII egress), eliminating third-party processor ingestion.
How does PrivacyScrubber establish Technical Safe Harbor under GDPR Article 32, CCPA/CPRA, and SOC 2 Type II?
Because free-text AI prompts can contain contextual narrative details, regulators evaluate compliance under risk minimization and state-of-the-art technical measures rather than blanket claims of regulatory non-applicability. PrivacyScrubber establishes a verifiable Technical Safe Harbor (Defense-in-Depth) under GDPR Article 32 (Recital 26 pseudonymisation), CCPA/CPRA, and SOC 2 Type II via: (1) 100% Direct PII Zero-Egress in local RAM, (2) Contextual Entropy Reduction across 30 Specialized Industry Profiles, and (3) Signed PDF Audit Receipts proving proactive statutory due diligence.
CISO AI Security Guide →