Canonical Python Reference Implementation of the Zero-Trust Data Sanitization (ZTDS) Protocol (IETF draft-sibiryakov-ztds-protocol-02, RFC v1.0).
ZTDS physically prevents PII, PHI, API tokens, and corporate secrets from crossing WAN sockets unmasked. It operates 100% in local process RAM with zero external network egress, sub-millisecond execution (<0.05ms), and verifiable memory zeroization.
- Invariant 1: Zero External Egress Prior to Sanitization Every byte of prompt data is inspected and sanitized locally in volatile memory. Zero unmasked data touches network interfaces.
-
Invariant 2: Deterministic Reversible Bijective Tokenization
Syntactic bracketed surrogates (
[NAME_1],[EMAIL_1],[API_SECRET_1]) preserve entity coreference and LLM reasoning semantics while enabling exact bijective restoration ($f^{-1}(f(x)) = x$ ). - Invariant 3: Ephemeral In-RAM Session Isolation & Zeroization (Theorem 2) Cryptographic mapping tables reside strictly in volatile RAM, isolated per session, and are destructively overwritten upon session completion.
- Invariant 4: Zero Subprocessors & Zero Telemetry No SaaS calls, no cloud APIs, and no telemetry side-channels. Eliminates GDPR Article 28 data processing agreements and vendor risk reviews.
Install the zero-dependency Python package:
pip install ztdsimport ztds
prompt = "Deploy microservice with key sk-ant-api03-0123456789abcdef0123456789abcdef and email dev@corp.internal"
# Use context manager for automatic Invariant 3 RAM zeroization
with ztds.session() as s:
# Phase 1: Local In-Memory Sanitization (<0.05ms)
sanitized, minted_tokens = s.sanitize(prompt)
print(sanitized)
# Output: "Deploy microservice with key [API_SECRET_1] and email [EMAIL_1]"
# Phase 2: Send sanitized string safely to any frontier model
# ai_raw_response = client.chat.completions.create(messages=[{"role": "user", "content": sanitized}])
ai_raw_response = "Confirmed. Credentials for [EMAIL_1] are stored."
# Phase 3: Bijective Restoration in Local RAM
clean_response = s.reveal(ai_raw_response)
print(clean_response)
# Output: "Confirmed. Credentials for dev@corp.internal are stored."
# Memory mapping is completely wiped upon context exit (Theorem 2)Install with LiteLLM extras:
pip install "ztds[litellm]"import litellm
from ztds.integrations import ZTDSGuardrail
# Intercept all calls locally before WAN socket dispatch
litellm.input_callbacks = [ZTDSGuardrail()]
litellm.output_callbacks = [ZTDSGuardrail()]
response = litellm.completion(
model="gpt-4o",
messages=[{"role": "user", "content": "My AWS key is AKIAIOSFODNN7EXAMPLE"}]
)
# Upstream OpenAI receives only "[API_SECRET_1]"
# Your application receives restored output automaticallyfrom langchain_openai import ChatOpenAI
from ztds.integrations import ZTDSSanitizingCallbackHandler
llm = ChatOpenAI(
callbacks=[ZTDSSanitizingCallbackHandler()]
)from crewai import Agent
from ztds.integrations import ZTDSSanitizerTool
security_agent = Agent(
role="Data Officer",
goal="Sanitize customer logs before delegation",
tools=[ZTDSSanitizerTool()]
)from llama_index.core import VectorStoreIndex
from ztds.integrations import ZTDSNodePostprocessor
query_engine = index.as_query_engine(
node_postprocessors=[ZTDSNodePostprocessor()]
)# Verify compliance with RFC v1.0 canonical test vectors
ztds conformance
# Output JSON for CI/CD pipelines
ztds conformance --json
# Audit local codebase for unmasked secrets and telemetry side-channels
ztds audit --dir ./src --strict
# Quick in-memory sanitization from command line
ztds sanitize "Contact alice@example.com with key sk-proj-12345678901234567890"All 8 canonical RFC v1.0 test vectors pass with 100% Invariant fidelity:
| Test Vector ID | Category | Status | Latency |
|---|---|---|---|
VEC-CORE-001 |
Universal Consumer PII | PASS | 0.04ms |
VEC-CORE-002 |
Coreference Determinism | PASS | 0.02ms |
VEC-HEALTH-001 |
Healthcare PHI & HIPAA Safe Harbor | PASS | 0.04ms |
VEC-FIN-001 |
Financial & Banking (ISO 20022) | PASS | 0.03ms |
VEC-SEC-001 |
Developer Secrets & Cloud Tokens | PASS | 0.03ms |
VEC-MULTI-001 |
Multi-Lingual Global Entities | PASS | 0.04ms |
VEC-REDOS-001 |
ReDoS Stress & Catastrophic Backtracking | PASS | 0.03ms |
VEC-UNICODE-001 |
Unicode Normalization & Homoglyphs | PASS | 0.03ms |
Average latency per prompt: ~0.031 ms (31 microseconds).
The open reference implementation (ztds) provides the universal baseline engine under the MIT License. For production corporate pipelines, regulatory safe harbors, and air-gapped enterprise workloads, refer to the Pioneer Certified Commercial Implementation:
- PrivacyScrubber Commercial SDK: Zero-egress in-memory Node.js & WASM engine engineered for ETL data lakes, RAG vector pipelines, and microservices (<0.05ms latency). Unlocks all 30 specialized industry profiles:
- Healthcare & PHI: HIPAA Safe Harbor (18 PHI identifiers, clinical MRNs, ICD-10)
- Financial & Banking: GLBA, ISO 20022 XML, SEPA pacs.008, SWIFT BIC, credit cards
- Legal & Dispute Resolution: FRE 502 (Attorney-Client privilege, work product doctrine, case dockets)
- Tax & Accounting: IRC § 7216 (W-2/1040 workpapers, EIN, PTIN, payroll ledgers)
- Defense & Aerospace: CMMC & ITAR (CUI tags, defense contract numbers, NATO codification)
- TEAMS & Enterprise Governance: Zero-knowledge cross-team session handoffs (Argon2id + XChaCha20-Poly1305), rootless air-gapped Docker sidecars, centralized regex governance, and DPO/CISO SHA-256 cryptographic compliance receipts.
- Zero-Trust MCP Server: Zero-Trust Data Sanitization stdio server for Claude Desktop, Cursor, and Windsurf (
uvx privacyscrubber-mcpornpx @privacyscrubber/mcp-server).
- IETF Specification: draft-sibiryakov-ztds-protocol-02
- Consortium & RFC Registry: https://ztds.ai
- IPR Disclosure: Filed under IETF RFC 8179 rules (Royalty-Free grant for Open-Source implementations with Defensive Termination).
- Commercial Certified Reference Implementation: PrivacyScrubber (30 specialized industry profiles, HIPAA/FRE 502/SOC 2 compliance packs, enterprise air-gapped on-premise engines).
MIT License. Copyright (c) 2026 Ilya Sibiryakov & ZTDS AI Consortium.