Home / Security
Five layers of enterprise-grade controls, from the disk your data lands on to the audit trail a compliance reviewer walks. Built in, not bolted on.
Three commitments
Not ours. Not OpenAI's. Not Anthropic's. AWS Bedrock terms explicitly prohibit customer data from training foundation models.
In-VPC Bedrock inference. Encrypted at rest with AWS KMS customer-managed keys. TLS 1.3 in transit between every hop.
Multi-tenant walls enforced in code. A continuous-integration lint gate rejects any pull request that would weaken the boundary.
Layer 5 · The proof
Governance, Audit & Compliance
Layer 4 · Your controls
Consent Governance & Session Privacy
Layer 3 · The AI boundary
AI Security & DLP
Layer 2 · The perimeter
Perimeter Security & IAM
Layer 1 · The foundation
Data Protection
Trust isn't a policy at Fingro. It's an architecture.
Trust doesn't come from any single control. It comes from every layer doing its job — and refusing to trust the layers around it.
Layer 1 assumes Layer 2 might fail. Layer 5 assumes all four below it might. Zero-trust, end-to-end.
Five layers · In-VPC AI · Deterministic DLP · SOC 1 & SOC 2 assessment in progress.
Layer 1 · The foundation
Layer 1 · The foundation
In practice
When a spouse joins your household, their account is walled off from your data by default. Sharing is an explicit action — never the state.
Layer 2 · The perimeter
In practice
Sign in with Face ID or a passkey. If you lose your device, you re-enroll from a trusted device or via MFA — no password to phish or reuse.
Layer 2 · The perimeter
Layer 3 · The AI boundary
Layer 3 · The AI boundary
In practice
When a couple asks WelX about filing jointly, that question — and every piece of context WelX pulls to answer it — stays inside AWS. Never OpenAI. Never Anthropic. Never leaves the VPC.
AI data path
Not every AI-powered fintech makes the same architectural choices. Here's what happens to your data across the three most common patterns — and where Fingro stands.
Dimension
Consumer AI
ChatGPT · Claude.ai · Gemini
Naive AI fintech
Third-party model calls
Fingro
In-VPC Bedrock
Where prompts go
OpenAI / Anthropic public API
Third-party model API
AWS Bedrock inside your VPC
Retention
Per vendor terms — often 30+ days
Per vendor terms — often opaque
Not retained by the model provider
Human review
Possible per vendor policy
Depends on vendor contract
Never
Used to train models
May be, per vendor opt-out state
Depends on vendor contract
Prohibited by Bedrock terms
PII redaction on I/O
None
Rare
SSN · EIN · card · routing (Luhn / ABA validated)
Layer 4 · Your controls
In practice
When a client shares 2024 tax documents with their CPA, the CPA sees exactly the 2024 return — not 2023, not the 2022 K-1. One tap revokes.
Layer 4 · Your controls
Layer 5 · The proof
Layer 5 · The proof
In practice
Every share granted, every share revoked, every admin action — written to an immutable audit stream that a compliance reviewer can trace end-to-end.
Operational
Versioned vendor list. Each bound by DPA to handle your data consistent with our Privacy Policy.
See the listEvery upload type-verified, script-stripped, malware-scanned by AWS-managed protection on arrival. Infected files quarantined.
Documented IR runbook. 72-hour breach notification obligation to affected users.
Vulnerability disclosure, vendor-review docs under NDA, or a question we haven't answered — security@fingro.ai.
See also: Privacy Policy · AI Disclosure · Data Processing Agreement