TrustBoost PII Sanitizer

A precision PII redaction layer for autonomous AI agent pipelines. Detects and redacts personally identifiable information before it reaches LLM providers, across English, Spanish (LATAM), Portuguese (BR/PT), German, Japanese, French, Italian, and Korean.

• Endpoint: https://api.trustboost.dev/sanitize
• Stack: FastAPI · OpenAI gpt-4o-mini (temperature 0) · Supabase · Solana payments via Helius

🛡️ Live Demo — no registration required

Try TrustBoost instantly in your browser: 👉 https://huggingface.co/spaces/TrustBoost/pii-sanitizer

Try it in 10 seconds — no wallet needed

curl -X POST https://api.trustboost.dev/sanitize/preview \
  -H "Content-Type: application/json" \
  -d '{"text": "My name is John Doe, email john@gmail.com, SSN 123-45-6789"}'

{
  "sanitized_content": "My name is [REDACTED], email [REDACTED], SSN [REDACTED]",
  "safety_score": 0.6,
  "risk_category": "PRIVATE",
  "demo": true,
  "requests_remaining": 2,
  "next": "https://github.com/teodorofodocrispin-cmyk/TrustBoost-PII-Sanitizer#trial"
}

3 free previews per IP · no account · no wallet · no setup. Ready for more? See Trial mode below — 50 free sanitizations with a Solana wallet.

Pay-per-call — /sanitize/quick (recommended for M2M agents)

The lightest-weight entry point for autonomous agents: pay $0.01 USDC per call, no TRIAL, no tx_hash, no prepaid bundle.

curl -X POST https://api.trustboost.dev/sanitize/quick \
  -H "Content-Type: application/json" \
  -H "PAYMENT-SIGNATURE: <base64-encoded x402 v2 signature>" \
  -d '{"text": "Contact John at john@company.com, SSN 123-45-6789"}'

• Networks: Base (eip155:8453, preferred) or Solana (solana:5eykt4...) — auto-detected
• Verified + settled via PayAI facilitator
• X-PAYMENT header also accepted (x402 v1, legacy)
• No text or no payment header yet? Returns a standard x402 v2 402 with payment instructions
• For 10,000+ calls, the prepaid bundle (149 USDC) is cheaper per call

MCP Server — Claude, Cursor & Windsurf native integration

TrustBoost is available as an MCP (Model Context Protocol) server. Add it to any MCP-compatible agent in one line:

{
  "mcpServers": {
    "trustboost": {
      "url": "https://api.trustboost.dev/mcp"
    }
  }
}

Once connected, your agent can call sanitize_pii automatically before sending any text to an LLM:

# Manifest
curl https://api.trustboost.dev/mcp

# Execute
curl -X POST https://api.trustboost.dev/mcp \
  -H "Content-Type: application/json" \
  -d '{"tool": "sanitize_pii", "input": {"text": "My email is john@gmail.com"}}'

Compatible with: Claude Code · Cursor · Windsurf · Any MCP-compatible agent

Quick start

curl -X POST https://api.trustboost.dev/sanitize \
  -H 'Content-Type: application/json' \
  -d '{
    "text": "My email is jane@example.com and my AWS key is AKIAIOSFODNN7EXAMPLE",
    "tx_hash": "TRIAL",
    "wallet_address": "your-agent-id"
  }'

Trial mode (tx_hash="TRIAL") gives 50 free sanitizations per wallet_address. Paid mode requires 149 USDC on Solana to the configured payment wallet, which unlocks 10,000 sanitizations per transaction signature.

Response schema (v2.6.0)

{
  "status": "success",
  "request_id": "TRIAL",
  "data": {
    "message": "Content successfully sanitized and logged.",
    "sanitized_content": "My email is [REDACTED] and my AWS key is [REDACTED]",
    "safety_score": 0.6,
    "risk_category": "CRITICAL",
    "entities_removed": true,
    "entities": [
      { "type": "email",          "category": "PRIVATE",  "redacted_text": "jane@example.com" },
      { "type": "aws_access_key", "category": "CRITICAL", "redacted_text": "AKIAIOSFODNN7EXAMPLE" }
    ],
    "redaction_source": "server",
    "timestamp": "2026-05-03T23:48:14.500705+00:00",
    "usage_metrics": { "quota_remaining": 48, "quota_limit": 50 }
  },
  "billing": { "license_type": "TRIAL", "status": "active" }
}

Field guide

Risk weights

safety_score is the sum of per-entity weights, capped at 1.0:

• CRITICAL → 0.40 (API keys, private keys, seed phrases, credentials, card numbers, …)
• PRIVATE → 0.20 (emails, phone numbers, national IDs, addresses, names, …)
• SENSITIVE→ 0.05 (handles, partial identifiers, DOB, …)

risk_category is the highest-severity tier with at least one entity, or "CLEAN" if entities is empty.

Server-side redaction enforcement (v2.6.0)

The model returns two things that have to agree: cleaned_text and entities. In practice they sometimes disagree — the model can correctly identify an entity in entities but fail to actually replace it in cleaned_text. That produces a sanitized_content that still leaks PII while the audit trail says everything is fine, which is worse than no audit trail.

v2.2 fixes this structurally. The model is now treated purely as a detector: it returns the entity list. The server is the redactor: for every entity whose redacted_text is a non-empty substring of the original input, the server replaces all occurrences with [REDACTED]. Long entities are processed before short ones to avoid partial overlap.

Conservative redaction by design: if the same value (e.g. 田中太郎) appears twice in the input, both occurrences are scrubbed.

The redaction_source field tells you what happened:

• "model" — the model's cleaned_text already matched the entity list, so server-side enforcement was a no-op (the model did its job).
• "server" — the server-side enforcer replaced one or more entities the model failed to remove. Track this metric over time as a model-reliability signal: a rising server rate means the prompt or model is drifting.
• "fallback_full_redaction" — the model returned malformed JSON; the failsafe parser triggered and the entire input was redacted as a single CRITICAL entity. Should be near-zero in steady state.

When the model's redacted_text does not appear verbatim in the input (paraphrasing, normalization, or hallucination), the entity is preserved in entities (and counts toward safety_score) but is also returned in unmatched_entities so callers can audit it.

Failure mode: fail-safe, not fail-open

If the upstream model returns malformed JSON, the response degrades to a single CRITICAL entity covering the entire input rather than risking a silent leak. Over-redaction is always preferred over under-redaction.

Languages and patterns

The system prompt covers, among others:

• English (global): emails, phones, SSN, credit cards, IBAN, IPs, addresses, and provider-specific API keys (OpenAI, Anthropic, GitHub, AWS, Google, Slack, HuggingFace, Stripe), private keys, crypto wallets, seed phrases.
• Spanish (LATAM): RFC, CURP, CUIT/CUIL, RUT, DNI, RUC, NIT, Cédula, country phones.
• Portuguese (BR/PT): CPF, CNPJ, RG, NIF, NUS, CEP, country phones.
• German (DE/AT/CH): Personalausweis, Steuer-IDs, Sozialversicherungsnummer, IBAN DE, addresses.
• Japanese: マイナンバー, 法人番号, 運転免許証, パスポート, 健康保険証, 電話番号, 住所, and 氏名 (full names in kanji, mixed scripts, katakana, or hiragana).
• French (FR/BE/CH/CA): NIR (Numéro de Sécurité Sociale), SIRET, SIREN, Carte Vitale, IBAN FR, Numéro fiscal, country phones +33/+32/+41.
• Italian (IT): Codice Fiscale, Partita IVA, Carta d'Identità (CIE), Tessera Sanitaria, IBAN IT, Patente di Guida, country phone +39.
• Korean (KR): 주민등록번호 (RRN), 사업자등록번호, 여권번호, 운전면허번호, 건강보험번호, 외국인등록번호, country phone +82.

Running locally

python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

cp .env.example .env  # then fill in real keys
uvicorn main:app --reload

Required environment variables:

• OPENAI_API_KEY
• SUPABASE_URL, SUPABASE_KEY
• HELIUS_API_KEY, PAYMENT_WALLET
• Optional: TRIAL_QUOTA (default 50), PAID_QUOTA (default 10000), REQUIRED_PAYMENT_USDC (default 149)

Tests

pip install pytest
python -m pytest tests/test_sanitize.py -v          # unit tests, no creds needed
TRUSTBOOST_LIVE=1 python -m pytest tests/test_live.py -v   # hits real /sanitize

The live tests consume TRIAL quota; set TRUSTBOOST_WALLET to a CI-specific identifier so they don't share quota with developer wallets.

Versioning