trustboost-api

TrustBoost PII Sanitizer is a precise PII (personally identifiable information) redaction layer for autonomous AI agent pipelines. Its core purpose is to detect and redact personally identifiable information before text reaches an LLM provider, preventing sensitive data leakage to the model. The product offers an online demo that can be tried in the browser without registration, and supports multilingual modes including English, Spanish (LATAM), Portuguese (BR/PT), German, Japanese, French, Italian, and Korean.

Main capabilities:

• Multilingual PII detection and redaction: covers emails, phone numbers, national identity documents, private keys, financial data, and more. The categories covered by the system prompt include: for English (Global), emails, phones, SSN, credit cards, IBAN, IP, addresses, as well as API keys, private keys, crypto wallets, and seed phrases for providers such as OpenAI, Anthropic, GitHub, AWS, Google, Slack, HuggingFace, and Stripe; for Spanish (LATAM), RFC, CURP, CUIT/CUIL, RUT, DNI, RUC, NIT, Cédula, and phone numbers of various countries; for Portuguese (BR/PT), CPF, CNPJ, RG, NIF, NUS, CEP, and phone numbers of various countries; for German (DE/AT/CH), Personalausweis, Steuer-IDs, Sozialversicherungsnummer, IBAN DE, and addresses; for 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, 氏名 (full names in kanji, mixed script, katakana, or hiragana); for French (FR/BE/CH/CA), NIR, SIRET, SIREN, Carte Vitale, IBAN FR, Numéro fiscal, and +33/+32/+41 phone numbers; for Italian, Codice Fiscale, Partita IVA, Carta d'Identità (CIE), Tessera Sanitaria, IBAN IT, Patente di Guida, and +39 phone numbers; for Korean, 주민등록번호 (RRN), 사업자등록번호, 여권번호, 운전면허번호, 건강보험번호, 외국인등록번호, and +82 phone numbers.
• Server-side redaction enforcement (v2.6.0): the model acts only as a detector and returns a list of entities, while the server acts as the redactorand performs the replacement. For each entity whose redacted_text is a non-empty substring of the original input, the server replaces all occurrences with [REDACTED], processing longer entities first to avoid partial overlaps. When the same value appears multiple times, all occurrences are removed. The redaction_source field indicates the execution status: "model" means the model's cleaned_text already matched the entity list and server-side enforcement was a no-op; "server" means the server replaced one or more entities that the model did not remove, which can be tracked as a model reliability signal; "fallback_full_redaction" means the model returned malformed JSON, triggering the fail-safe parser to redact the entire input as a single CRITICAL entity, and should be close to zero in steady state. When the model's redacted_text does not appear verbatim in the input, the entity remains in entities and counts toward safety_score, while also being returned in unmatched_entities for caller auditing.
• Failure mode: fail-safe rather than 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 preferable to under-redaction.
• Response schema (v2.6.0): safety_score is the sum of the weights of each entity, capped at 1.0. CRITICAL → 0.40 (API keys, private keys, seed phrases, credentials, card numbers, etc.); PRIVATE → 0.20 (emails, phones, national IDs, addresses, names, etc.); SENSITIVE → 0.05 (account numbers, some identifiers, dates of birth, etc.). risk_category is the highest severity level among at least one entity, or "CLEAN" if entities is empty.

Usage:

• Endpoint:https://api.trustboost.dev/sanitize
• Tech stack: FastAPI · OpenAI gpt-4o-mini (temperature 0) · Supabase · Solana payments via Helius
• Online demo:https://huggingface.co/spaces/TrustBoost/pii-sanitizer（每个 IP 3 free previews, no account, wallet, or setup required)
• Pay-per-call /sanitize/quick: the lightest entry point for autonomous agents, $0.01 USDC per call, no TRIAL, tx_hash, or prepaid package required. Networks: Base (eip155:8453, preferred) or Solana (solana:5eykt4...), auto-detected. Verified and settled via PayAI facilitator. Also accepts X-PAYMENT header (x402 v1, legacy). Returns a standard x402 v2 402 with payment instructions when there is no text or no payment header. At 10,000+ calls, the prepaid package (149 USDC) is cheaper per call.
• MCP server: compatible with Claude Code, Cursor, Windsurf, and any MCP-compatible agent, can be integrated in one line. Once connected, the agent can automatically call sanitize_pii before sending any text to the LLM.
• Trial mode (tx_hash="TRIAL"): 50 free redactions per wallet_address. Paid mode requires paying 149 USDC on Solana to the configured payment wallet; each transaction signature unlocks 10,000 redactions.
• Environment variables required to run locally: SUPABASE_URL, SUPABASE_KEY, HELIUS_API_KEY, PAYMENT_WALLET. Optional: TRIAL_QUOTA (default 50), PAID_QUOTA (default 10000), REQUIRED_PAYMENT_USDC (default 149). Live tests consume TRIAL quota; TRUSTBOOST_WALLET can be set to a CI-specific identifier to avoid sharing quota with the developer wallet.

Version evolution:

• 2.6: Redaction proof on Solana via Helius. Each paid redaction is anchored on-chain and can be verified at /verify/{anchor_tx}. x402 native, HTTP 402 and autonomous payment instructions. CORS for browser agents. agent-card.json for Circle Agent Stack discovery.
• 2.5: Per-agent privacy budget, M2M TrustBoost Score.
• 2.4: Per-agent privacy budget, agent_budgets table in Supabase. Operators configure daily limits, and agents operate autonomously within the limits.
• 2.5: TrustBoost Score: /score/{wallet} endpoint, M2M trust verification, trust tiers (TRUSTED/VERIFIED/ACTIVE/NEW), aggregated from audit_log.
• 2.3: Context-aware redaction: context field in /sanitize (legal/financial/medical/code/general), adjusts redaction depth by context type, adds context_applied in the response.
• 2.2: Server-side redaction enforcer, redaction_source telemetry, unmatched_entities audit field.
• 2.1: Structured entities array, server-side deterministic scoring, strengthened JSON parsing, improved Japanese name detection.
• 2.0: Multilingual prompt rewriting.

Use cases: scenarios in autonomous AI agent pipelines where personally identifiable information must be redacted for compliance before text is sent to an LLM, especially suitable for multilingual environments and M2M applications that require audit trails and on-chain proof.

Limitations: The documentation does not provide information about price changes, model accuracy guarantees, compliance certification for specific jurisdictions, or service availability SLAs.