Trust

CrossCode — Trust The information below is intended for the IT, security, and compliance leads at organizations evaluating CrossCode. Last updated: 2026-06-22.

How CrossCode handles your trade data

CrossCode is an AI-assisted customs classification platform used by licensed customs brokers and importer trade compliance teams. We work with commercial invoices, line-item descriptions, supplier identities, and other sensitive trade data. We have built CrossCode around four commitments.

1. Your data is yours. Submissions and classification artifacts are Customer Data. You own them. We hold them under contract to operate the Services for you, not for any other purpose.

2. We do not train on your invoices. We do not use Customer Data — including any Submission, Customer-identifiable Output, or Customer-identifiable record — to train, retrain, fine-tune, evaluate, benchmark, or otherwise improve any AI or machine-learning model (including any foundation model, classifier, retrieval index, or evaluation suite). We do not opt you in. If we ever want to use a customer's data in a way that would change this, we ask in writing, and you can decline without affecting your service. We configure our LLM inference providers (Anthropic, OpenAI) to minimize data retention and prohibit training use to the extent their API tiers allow, and we maintain documentation of that configuration. See our Sub-Processor List for the specific data handling terms that apply to each provider.

3. We are a decision-support tool, not a customs broker. Output is a recommendation. The licensed broker and/or importer of record remains responsible for every classification decision. We support your duty of Reasonable Care under 19 U.S.C. § 1484; we do not transact Customs Business under 19 C.F.R. § 111.1.

4. We retain classification artifacts for as long as you need them. By default, we retain the Submission, Output, model identifier and configuration, prompt template hash, retrieval index hash, and cited rulings for the longest of your subscription, your elected retention window, five years from the date we generated the classification, or five years from the date of the corresponding customs entry where you supply that date. The CBP recordkeeping clock under 19 U.S.C. § 1508(c) runs from the date of entry — please confirm our retention covers your actual entry dates. We are not your recordkeeper; we are a defensible companion to your records.

Security at a glance

Control	Detail
Encryption in transit	TLS 1.2 or higher
Encryption at rest	AES-256
Authentication	MFA required for all employee production access; SSO available for customers
Access control	Role-based, least-privilege; quarterly access reviews
Logging	Centralized access logs retained 24 months
Backups	Daily, encrypted, separate region
Vulnerability management	Monthly scans; patch SLAs: critical 7 days, high 30 days
Penetration testing	Annual third-party pen test (initial: within 12 months of go-live)
Personnel	Background checks, NDAs, annual security and AI safety training
Incident response	Documented runbook; 72-hour breach notification

Frameworks and certifications

Aligned today with the NIST AI Risk Management Framework (NIST AI 100-1) and ISO/IEC 27001 control families.
In flight: SOC 2 Type I — target completion within 12 months of 2026-06-22.
Next: SOC 2 Type II — target within 24 months. ISO/IEC 42001 alignment as the AI management-system standard matures into customer demand.

We do not claim certifications we do not hold. When the SOC 2 reports are available, we share them under NDA on request.

AI safety practices

Versioned model releases. Every Output records the model identifier and configuration, prompt template hash, and retrieval index hash used to produce it, so the inputs and configuration behind any classification you've run with us remain auditable for the full retention period. Foundation-model providers retire model versions over time; we preserve the recorded identifier and any contemporaneous evaluation results — we do not warrant the runtime reproducibility of provider-retired models.
Multi-provider LLM architecture. We use Anthropic and OpenAI as inference providers and route between them. We configure provider API settings to minimize data retention and prohibit training use to the extent available under each provider's current API tier, and we preserve evidence of that configuration. See our Sub-Processor List for the specific retention terms that apply to each provider.
Pre-release evaluation. New model versions are evaluated against a held-out classification benchmark, including CROSS-ruling alignment and Section 301/232 awareness, before promotion to production.
Prompt-injection hardening. Submissions are treated as untrusted input. We sanitize and segregate Submission content from agent instructions.

Sub-processors

A current list of Sub-processors is at crosscode.pro/subprocessors. Material additions are notified at least 30 days in advance to customers who subscribe to updates.

Data residency

CrossCode is hosted in the United States today. EU data residency is on the roadmap and is gated by customer demand. If you require EU-only processing, please contact us.

How to ask us a hard question

Email trust@crosscode.pro. Real human responses, usually within a business day. For security disclosures, see security@crosscode.pro.

CrossCode, Inc. (a Delaware corporation in formation) — operating as CrossCode.