Nesta página
Fhinck Trust Center
How we protect your operation's data: controls, evidence, independent assessments and documents, in one place.
Last updated: October 2026
Security Overview
Fhinck adopts a Defense in Depth approach, implementing multiple layers of security controls across our entire infrastructure, applications, and operational processes.
Zero Trust
Privacy by Design
Data Minimization
Layered Defense
Transparency
Cloud Infrastructure
Secure and certified hosting
Google Cloud Platform (GCP)
Region: Google Cloud, primary processing in the USA
GCP Certifications:
- ISO 27001 / 27017 / 27018: Information security management, cloud security, and personal data protection
- SOC 2 Type II: Security, availability, and confidentiality controls — independently audited
- PCI DSS Level 1: International standard for payment data protection
- FedRAMP / C5 / HIPAA: US Government, German BSI, and healthcare sector certifications
- GDPR/LGPD Compliant: Data protection clauses (LGPD — Brazilian Data Protection Authority)
SLA & Availability:
GCP SLA: 99,95%
Fhinck Commitment: 99,5%
The 99.5% commitment is deliberately conservative: it sits below GCP's 99.95% infrastructure SLA because it reserves headroom for the application layer — deploys and maintenance windows — instead of passing the cloud's ceiling off as the whole platform's.
Serverless Architecture:
- Cloud Functions & Cloud Run
- Automatic scaling
- Automatic security patches
Secure Databases
BigQuery (Data Warehouse):
- • AES-256 encryption at rest
- • Granular access controls
- • Complete query auditing
Firestore (NoSQL):
- • Automatic encryption
- • Custom security rules
Environment Segregation
Development
Synthetic or anonymized data
Staging
Isolated pre-production testing
Production
Protected and monitored environment
Encryption
Encryption in Transit
TLS 1.2+ (1.3 preferred)
- All communications via HTTPS
- SSL/TLS certificates managed and renewed automatically — no manual intervention, no expiry risk, continuously checked
- Hardened configuration (no weak ciphers)
- HSTS (HTTP Strict Transport Security)
Client → TLS 1.3 → Load Balancer → Internal TLS → Backend
Encryption at Rest
AES-256 (Advanced Encryption Standard)
- Data stored encrypted in BigQuery
- Recovery copies encrypted at rest (AES-256, managed by GCP)
Key Management:
- • Secrets in Google Secret Manager, with automated secret scanning of the code base
- • At-rest encryption managed by GCP (AES-256, Google keys)
- • Per-client isolation: dedicated BigQuery dataset
Password Hashing:
Firebase Auth (scrypt) - managed password hashing, resistant to brute-force attacks
Pseudonymization and Anonymization
Options available to clients:
Pseudonymization
ID replaced by unique hash
Full Anonymization
Removal of personal identifiers
Aggregation
Aggregated data only
Access Controls and Authentication
Multi-Factor Authentication (MFA)
Mandatory for all critical access:
Single Sign-On (SSO)
Available Integrations:
- SAML 2.0: Enterprise standard
- OIDC: Modern identity protocol
- Google, Azure AD, Okta, Auth0
Benefits:
- Centralized identity management
- Single credentials for multiple systems
- Corporate password policies enforced
- Instant access revocation
Role-Based Access Control (RBAC)
Granular profiles in the Dashboard:
Viewer
Read-only access
Analyst
Report creation
Manager
Department access
Administrator
Full management
Auditor
Log access
Principle of Least Privilege: Each user receives only the permissions required for their role.
Secure Software Development
SAST
Static Code Analysis
- • ESLint Security Plugin
- • Semgrep (OWASP/CWE)
- • Gitleaks (secrets)
DAST
Dynamic Application Analysis
- • OWASP ZAP
- • Baseline scan (passive), weekly
Vulnerabilities
Proactive Management
- • npm audit (daily)
- • Trivy (daily: dependencies and infrastructure as code)
- • OWASP ZAP (weekly: dynamic scan)
Vulnerability Remediation Timelines
CRITICAL
< 48h
HIGH
< 10 days
MEDIUM
< 30 days
LOW
< 90 days
Web Security Assessment
Frequency
Monthly
Last Performed
September 2026
Result
✓ No critical vulnerabilities
Privacy and Data Protection
LGPD Compliance
Status: Aligned with the LGPD
Roles and Responsibilities
Controller: Client
Determines the purposes and means of processing
Processor: Fhinck
Processes data according to Controller instructions
DPO: Michel Zarzour Filho
dpo@fhinck.com
(11) 98367-3803
Data Minimization
✓ We Collect
- • User ID, hostname
- • App metadata
- • Activity timestamps
✗ We Do NOT Collect
- • Background screenshots
- • File reading
Security Incident Response
Response Process
Detection
0–2 hours
Containment
2–8 hours
Notification
Within 48h
Resolution
Post-incident
Emergency Response Team (ERT)
Audits and Compliance
Internal Audits
Compliance and references
LGPD Compliance
Aligned
Marco Civil da Internet
Compliant
OWASP Top 10
Mitigations implemented
Why Continuous Compliance
Security that evolves with the operation — not an annual snapshot
Continuous controls, not once a year
Static security analysis and dependency checks run every day, and the dynamic scan runs every week, all year long. Always-current evidence is worth more than an attestation issued on a single day.
AI-first: operations change constantly
With AI, flows and architecture evolve continuously. A certification locked to a frozen scope becomes stale within months. Our governance keeps pace with change instead of constraining it.
Direct investment in real protection
Instead of directing budget into the audit/certification cycle, Fhinck invests directly in controls, real-time monitoring, and an automated monthly web security assessment (Beagle Security) — protection where it creates real results.
Security Controls Matrix
Structured view of the 10 control domains and how each one operates today
Security for an AI-first operation
The frameworks above, SOC 2 and ISO/IEC 27001, describe controls designed for a human-run operation, and Fhinck is aligned with them. But Fhinck runs on AI agents. For an AI-first platform, the parameters that actually matter are the ones defined by AI-specific standards: the NIST AI Risk Management Framework, ISO/IEC 42001 (AI management), and the OWASP Top 10 for LLM and for Agentic Applications. Below are the risks those frameworks classify as critical for AI systems — and how Fhinck addresses each, with evidence in its own repository.
The AI-first lens
A traditional audit asks “is a human approving this?”. In an agentic operation the right question is “what stops the agent from acting on its own — and is that testable?”.
Prompt injection and jailbreak defense
Active defenseReference standard: OWASP LLM01:2025 · Agentic ASI01
Any input reaching a model may try to hijack its instructions. We treat user input as untrusted.
How Fhinck delivers it
Dashboard copilot: deterministic detection of injection and jailbreak in Portuguese and English (instruction override, role-hijack, system-prompt extraction attempts, token smuggling), a 4,000-character cap, HTML/script stripping, and a dedicated per-session adversarial rate limit. The backend pipeline applies its own defense (sandwich defense).
Output guardrails and leak prevention
Active defenseReference standard: OWASP LLM02 · LLM05 · LLM07
No AI response leaves without being filtered against leaking personal data, another client's data, or the system's own.
How Fhinck delivers it
The output filter redacts CPF, CNPJ, external email, card, phone and tokens; a canary token discards the whole response if there is a sign of system-prompt extraction. In the support agent, 15 deterministic leak patterns (client datasets, internal IDs, table and workflow names) halt execution.
Agent harness and behavioral contract
Hook-enforcedReference standard: OWASP LLM06 · Agentic ASI02/ASI03
The #1 risk of agentic systems is “excessive agency”. Each agent runs under a mechanical contract, not on trust in the model.
How Fhinck delivers it
Subagents only run when: declared skills were actually invoked, quantitative claims carry structural evidence (file:line or output), and a persistent artifact exists — rules C1–C4 enforced by a hook that blocks the call. A briefing-size limit is enforced the same way, by a hook that blocks the call.
Human approval for critical actions
Hook-enforcedReference standard: OWASP Agentic ASI09 · NIST AI RMF
A destructive or production action never happens just because the agent decided to.
How Fhinck delivers it
Critical commands (production deploy, DROP/TRUNCATE, force-push, cloud delete) require a one-time code (TOTP) + an offline-verified cryptographic signature, or a same-day Triple Gate artifact. Freeze windows block deploys during client hours, and every bypass produces an audit trail.
Continuous adversarial red-teaming and LLM-as-judge
ContinuousReference standard: NIST AI RMF (MEASURE) · MITRE ATLAS
Agents are attacked on purpose and judged on quality before any promotion.
How Fhinck delivers it
Internal agent-audit framework with 33 adversarial cases (prompt injection, confirmation bypass, permission abuse). Support-pipeline changes only ship after an LLM-as-judge gate across 4 dimensions (action, quality, compliance, effectiveness) with a minimum 4.5/5, over a synthetic suite plus sanitized real tickets.
Observability of every agent run
In productionReference standard: OWASP Agentic ASI10 · NIST AI RMF
An agent's decision can't be a black box: it has to be auditable afterward.
How Fhinck delivers it
A single OTel-GenAI-shaped telemetry channel in BigQuery (hook decisions, agent runs, deploys, audit findings), with automatic PII redaction and 365-day retention; an append-only skill-invocation log used for enforcement.
Tenant isolation and contained blast radius
Structural isolationReference standard: OWASP LLM02:2025 · NIST AI RMF (MANAGE)
In an agentic operation, each client's data stays in its own dataset and every query is pinned to a single client, which limits the blast radius of a failure.
How Fhinck delivers it
Each client lives in a dedicated BigQuery dataset, and every query is scoped and sanitized per company before it touches data — the company identifier is validated and pinned into the table path, with no cross-client queries. Service credentials follow least privilege.
AI-First control equivalence
The same objective behind every control, met in an AI-native way
Why we don't pursue the badge
SOC 2 and ISO 27001 were written for a company where every control is performed and approved by a person. Fhinck operates through AI: the badge would certify human procedures that don't describe how we actually operate. We meet the objective of every control — we just don't perform the activity in the human form the standard presumes.
Objective vs. activity
Every standard has two layers: the control objective (what it protects) and the prescribed activity (the human procedure). Fhinck delivers every objective through machine mechanisms — auditable and continuous.
What the standard presumes — human operation
How Fhinck meets it — AI operation
A.6.1 · Background screening
What the standard presumes — human operation
HR investigates the history of a hired person.
How Fhinck meets it — AI operation
Model evaluation and red-teaming + behavior fixed in a versioned, auditable system prompt.
A.6.3 · Training and awareness
What the standard presumes — human operation
Employee takes periodic security training.
How Fhinck meets it — AI operation
Guardrails and input sanitization against prompt injection — safe behavior is coded, not trained person by person.
CC5 · Segregation of duties
What the standard presumes — human operation
Two different people: no one approves their own work.
How Fhinck meets it — AI operation
Automated gates + an autonomy × risk model: higher-impact actions don’t execute without independent verification.
CC6 / A.5.15–18 · Identity-based access
What the standard presumes — human operation
Human user with MFA; provisioning and offboarding by HR.
How Fhinck meets it — AI operation
Machine identity with a named owner, least privilege, short-lived access tokens and lifecycle decommissioning.
CC8 / A.8.32 · Change management
What the standard presumes — human operation
A second human reviews and approves every change.
How Fhinck meets it — AI operation
Freeze + snapshot + rollback + automated gate at every deploy, with a full audit trail.
Processing Integrity
What the standard presumes — human operation
Deterministic process, verifiable step by step.
How Fhinck meets it — AI operation
Mandatory human-in-the-loop for side-effect actions + audit log of AI decisions.
CC7 / A.8.15 · Operations and logging
What the standard presumes — human operation
On-call analyst monitors and records events.
How Fhinck meets it — AI operation
Real-time observability (Cloud Logging + BigQuery), machine-auditable and always current.
What the badge doesn't even ask
Where AI operation introduces risks these standards — predating the agentic era — don't even address (autonomy and blast radius, prompt injection, machine identities), we apply specific governance and OWASP LLM Top 10 mapping. Passing a pre-agentic audit doesn't cover these risks; we do.
For security and procurement teams
The question that matters in a vendor assessment isn’t "do you have the badge?" — it’s "can you prove your controls?". We answer in two ways:
Infrastructure already certified by a third party
Everything runs on Google Cloud Platform — SOC 2 Type II and ISO 27001, independently audited. The infrastructure layer you need to validate already has formal attestation.
We answer your questionnaire
We complete vendor security questionnaires (SIG, CAIQ or your format) directly, control by control.
SOC 214
Fhinck is aligned with this framework’s controls, but does NOT hold and is not pursuing SOC 2 certification.
Fhinck’s security controls are aligned with the SOC 2 Trust Services Criteria. Our underlying infrastructure (Google Cloud Platform) is SOC 2 Type II certified. Honest status per criterion and per Common Criteria below.
ISO 2700113
Fhinck is aligned with this framework’s controls, but does NOT hold and is not pursuing ISO 27001 certification.
Fhinck’s technical and organizational controls are aligned with Annex A of ISO/IEC 27001:2022 (93 controls across 4 themes). Honest status per theme and per representative control below.
Documents
Public documents can be downloaded directly. Documents marked as restricted contain internal information: request access with your corporate email and receive a one-time code.
Consolidated Institutional Policies and Practices (CPP)
PDF · ~11.5 MB
Master document gathering the Information Security Policy, access management, cryptography, incident response, business continuity and the privacy program (LGPD). Revised version with formal version control.
Internal Audit Report — October/2025
PDF · ~373 KB
Results of the internal security and privacy audit conducted in October 2025, covering access controls, cryptography, continuity and data handling.
Mutual Non-Disclosure Agreement (NDA)
PDF · ~96 KB
Fhinck's mutual NDA template for exchanging confidential information during evaluation and service delivery. Base document for signature between the parties.
Subprocessors
Vendors that process data on behalf of Fhinck, the same ones declared in our Privacy Policy.
| Vendor | Purpose | Location |
|---|---|---|
| Google Cloud Platform | Cloud infrastructure: processing and storage of platform data | United States (primary region) |
| Generative AI providers | Language models used in the platform's AI features, under contractual safeguards | — |
| Analytics and observability tools | Observability and platform usage analytics | — |
| Google reCAPTCHA Enterprise | Protection against automated submissions on website forms | — |
The named subprocessor list, with purpose and country, can be requested from our DPO at dpo@fhinck.com.
Security updates
What recently changed in our security posture.
Oct 2026
Trust Center open to the public
Controls, evidence and independent assessments are now open to any visitor. Documents with internal information remain available on request.
Sep 2026
External attack surface hardened
DNS record defining which authorities may issue certificates for fhinck.com, SSH access to automation machines restricted to an authenticated tunnel, and a review of DNS records. On Sep 28, SecurityScorecard recorded an A rating, with 100 on all 10 factors assessed.
Sep 2026
Automated web security assessment
Monthly Beagle Security run on Sep 11: risk rated low and no critical vulnerabilities.
Jun 2026
Controls mapped to SOC 2 and ISO 27001
Per-criterion status published on this page, with evidence for each control. Fhinck is aligned with these standards, without certification.
Responsible vulnerability disclosure
Found a possible vulnerability in a Fhinck system? We want to hear about it. Good-faith research that follows this policy is welcome.
How to report
- 1Email security@fhinck.com with a description, steps to reproduce and the impact you observed.
- 2We acknowledge receipt within 48 hours and keep you informed until the fix.
- 3Please keep the finding confidential until it is fixed or for 90 days, whichever comes first.
In scope
fhinck.com and its subdomains, the platform dashboard and API.
Out of scope
Denial of service, social engineering, physical access, spam, and any access, modification or retention of customer data beyond the minimum needed to demonstrate the issue.
Safe harbor
Good-faith research conducted within this policy is considered authorized, and Fhinck will not take legal action against it. There is no monetary bounty program.
The same contact is published at fhinck.com/.well-known/security.txt, following RFC 9116.
Contact and Support
Emergencies
ERT Team
24/7 for critical incidents
Our Commitment
The security and privacy of our clients' and users' data are absolute priorities at Fhinck. We continuously invest in technology, processes, and people to maintain the highest standards of protection.
Last updated: October 2026