Security Whitepaper
Orange AI Security Policy
This document details how the Orange AI enterprise customer service platform protects your business data and customer privacy at the architectural level. We apply industry-leading security practices to ensure every enterprise customer's data remains safe and secure.
Version: v1.2 ¡ Last Updated: September 2026
|
1. Security Commitment
Orange AI recognizes that enterprise customers are entrusting us with their most sensitive business knowledge and customer conversation data. This trust is our most important responsibility. Security is not an afterthought â it is the first principle of every system design decision.
We commit to the following core principles:
Defense in Depth
Every layer of data access has independent security validation, from the frontend Widget to backend Cloud Functions.
Principle of Least Privilege
Each role is granted only the minimum permissions required to perform its function. Over-provisioning is strictly avoided.
Full Auditability
All critical operations are recorded in audit logs, giving enterprise administrators complete visibility into system usage at any time.
Data Sovereignty
Your business data belongs entirely to your organization. Orange AI does not share data across tenants or use your data for model training.
2. Infrastructure Security
Orange AI is built entirely on Google Cloud Platform (GCP), inheriting Google's world-class infrastructure security guarantees.
Industry-Leading Cloud Security Certifications
Our cloud infrastructure holds ISO 27001, SOC 1, SOC 2, and SOC 3 certifications, with comprehensive GDPR and HIPAA compliance support. Google data centers adhere to the most rigorous physical and digital security standards in the industry.
Core Infrastructure Components
Distributed Enterprise Cloud Database
The primary enterprise data store. All records are encrypted at rest with AES-256 and protected by strict server-side access control rules.
Serverless Microservices
All AI inference scheduling, vector searches, and sensitive operations run in isolated serverless sandboxes. The frontend never touches sensitive business logic or API keys.
Enterprise Encrypted Object Storage
Knowledge documents and rich media files are logically partitioned and isolated per tenant, backed by strict server-side rules preventing cross-organization access.
Enterprise Identity & Access Management (IAM)
Google-managed enterprise authentication supporting Google/Apple SSO and enterprise email/password login, with zero plaintext credential storage.
Data Residency & Infrastructure in Taiwan
Orange AI places the highest priority on data sovereignty and regional compliance. All core computing nodes and primary production databases are strictly hosted within the Google Cloud Taiwan Data Center, ensuring complete Data at Rest in Taiwan compliance.
Distributed Cloud Database
Region: Taiwan Data Center
All conversation logs, organization records, and configurations encrypted locally
Serverless Microservices
Region: Taiwan Compute Nodes
RAG retrieval orchestrations and channel webhooks execute in local serverless sandboxes
Encrypted Object Storage
Region: Taiwan Storage Space
Uploaded knowledge source files and conversation attachments remain entirely within Taiwan
shield_with_heart Enterprise Cross-Region Disaster Recovery Vault
To meet enterprise Business Continuity Planning (BCP) standards against severe natural disasters or disruptions, encrypted database backups are automatically mirrored daily to an isolated Asia-Pacific Cross-Region Disaster Recovery Vault project. The vault is guarded by a 30-day immutability retention policy, ensuring reliable multi-region disaster recovery (RPO ≤ 24h, RTO ≤ 4h).
3. Authentication
Orange AI uses a dual-track authentication design, providing appropriate security levels for the enterprise management console and the external visitor Widget respectively.
Enterprise Console
- check_circleGoogle / Apple OAuth 2.0 social login
- check_circleEmail + password authentication
- check_circleSecure JWT identity token validation throughout
- check_circleCustom role claims injection (RBAC)
- check_circleAutomatic token rotation and expiry
External Visitor Widget
- check_circleSecure anonymous session authentication
- check_circleAccess restricted to own conversation only
- check_circleCannot access org settings or knowledge base
- check_circleRate limiting to prevent abuse
- check_circleAPI Token embedded authentication
4. Role-Based Access Control (RBAC)
The system implements RBAC using server-side custom role claims, embedding role information directly within secure JWT tokens. This ensures that even if application logic is bypassed, server-side access control rules and backend microservices independently enforce access control.
| Role | Access Scope | Restrictions |
|---|---|---|
| Super Admin | Platform-wide management (Orange AI team only) | Manages organizations via Admin Console; does not access customer business data |
| Org Admin | Full org settings, member management, knowledge base, Agent config, tool integrations | Limited to own organization; cannot access other orgs |
| Knowledge Manager | Maintain knowledge base, adjust Agent personas, configure external tools, read-only conversation access | Cannot reply to or take over real user conversations; no access to member management, org settings, or audit logs |
| Agent Customer Service |
Reply to and take over real user conversations | Cannot modify knowledge base, Agent settings, or access any org management functions |
| External Visitor | Chat via Widget (anonymous) | Own conversation only; no internal org access whatsoever |
Permission Matrix
| Feature | Admin | Knowledge Manager | Agent |
|---|---|---|---|
| Knowledge base management (add/edit/delete docs) | â | â | â |
| Agent persona settings (edit system prompts) | â | â | â |
| Tool integrations (add/edit external tools) | â | â | â |
| Conversation testing (RAG simulator) | â | â | â |
| Reply to real conversations (Inbox takeover) | â | â | â |
| Org settings (billing, quota, system config) | â | â | â |
| Audit logs (view) | â | â | â |
| Member management (invite/remove) | â | â | â |
Knowledge Manager Conversation Isolation
Knowledge Managers may read conversations in a read-only capacity to help optimize the knowledge base. The frontend hides takeover buttons and input fields for this role, while backend access control rules and microservices independently block any write operations to real user conversations â ensuring dual-layer enforcement.
Frontend Route Guards & State Validation
The frontend enforces client-side route guards combined with reactive state validation to prevent unauthorized users from navigating to protected views. Backend Security Rules serve as the independent final line of defense â even if the frontend interface is bypassed, data access is strictly blocked.
5. Multi-Tenant Data Isolation
Orange AI employs strict multi-tenant isolation architecture, ensuring that data from different enterprises is separated at the architectural level â not merely by application logic.
Database Access Control
All database reads and writes are governed by server-side security rules. Every request must carry a valid authentication credential, and the requester's organizational membership is verified server-side â ensuring that no data belonging to another organization can be accessed, even if application-layer logic is bypassed. The backend rules act as an independent final line of defense.
System logs and audit records carry additional write protection, allowing only trusted backend services to write. Any write attempt from the frontend or a general user is rejected at the rule layer.
File Storage Isolation
All organization-related files are partitioned by organizational membership. Storage-layer access rules ensure that cross-organization access requests are rejected server-side, independent of any application-layer logic.
Vector Knowledge Base Isolation
Each organization's vector data (Embeddings) is stored in an isolated, dedicated space. RAG queries are strictly scoped on the backend to the requesting organization's vector space. Cross-tenant knowledge base access is architecturally impossible â not merely blocked by application logic.
6. Encryption & Transport Security
Encryption in Transit
- check_circleAll API communication enforces HTTPS / TLS 1.3
- check_circleStreaming responses use WSS encryption
- check_circleHSTS headers enforce secure connections
Encryption at Rest
- check_circleAll database data encrypted at rest with AES-256
- check_circleObject storage files automatically encrypted
- check_circleKeys managed by Google Cloud KMS
7. Enterprise Data Inventory & End-to-End Data Flow
Orange AI implements rigorous data governance and asset inventory standards. To provide enterprise legal, security, and audit teams with full transparency, the matrix below outlines all enterprise and customer data assets accessed, stored, and processed by our platform.
Enterprise Data Inventory Matrix
| Category | Content & Sources | Storage & Encryption | Lifecycle & Retention | AI Model Training Policy |
|---|---|---|---|---|
| description Knowledge Docs | Internal regulations, product FAQs, user manuals, uploaded PDF, DOCX, TXT, CSV files | Encrypted object storage (tenant-partitioned, AES-256 encrypted); semantic embeddings stored in dedicated vector collections | Managed by enterprise admins; purged permanently upon document deletion or org termination | Strictly prohibited from training public foundation models (RAG retrieval only) |
| language Website & Web Content | Authorized crawled public official websites, service terms, and online FAQ text | Encrypted object storage (Markdown files) & dedicated org vector storage | Synchronized with knowledge base; can be recrawled or deleted anytime | Strictly prohibited from public model training |
| chat Omnichannel Conversations | Customer messages across LINE, Facebook, Instagram, WhatsApp, and Web Widget (text, voice transcripts, images, attachments) | Chat records stored in dedicated database (AES-256 encrypted); multimedia attachments stored in encrypted object storage (10MB limit) | Text stored in org Inbox; attachments automatically deleted 30 days after chat archive | Gemini Enterprise API Zero Data Retention & Training Restrictions guaranteed |
| account_box Customer Profiles | Channel anonymous IDs (LINE userId, Meta PSID, etc.), authorized public display names, and avatars | Tenant-isolated conversation store in dedicated database, encrypted at rest | Channel avatars/names cached with 30-day TTL; purged upon organization termination | Never used for AI model training |
| database Custom DB & Bookings (OMS) | If Order Management System (OMS) is enabled: booking form submissions, names, phone numbers, order items | Tenant-dedicated custom database store | Automatically cleaned per enterprise-configured TTL retention schedules; exportable | Never used for AI model training |
| badge Admin & Staff Accounts | Admin/agent login emails, full names, IAM UIDs, RBAC role claims | Enterprise Identity and Access Management (Google-managed, hashed credentials, no plaintext) & dedicated database | Revoked immediately upon account deactivation or member removal | Not Applicable (Identity Data) |
| history_edu Audit & Activity Logs | Actor ID, millisecond timestamps, action categories (Prompt tuning, doc upload, integrations), field diffs | Dedicated audit log repository, writable only by serverless microservices, tamper-proof | Retained per enterprise compliance requirements; exportable as CSV | Not Applicable (Security Log) |
| vpn_key Integration Credentials | LINE Channel Secret/Token, Meta tokens, Google Workspace OAuth Refresh Tokens | Google Cloud Secret Manager / Server-side isolated database records, strictly hidden from frontend | Deleted immediately upon channel or integration disconnection | Not Applicable (API Credentials) |
End-to-End Data Flow & Privacy Safeguards
Secure Inbound Transmission & Signature Verification
Customer inputs sent via LINE or the Web Widget travel over HTTPS / TLS 1.3 directly to secure microservice endpoints in the Google Cloud Taiwan Data Center. Cryptographic signatures (such as LINE's x-line-signature) are verified on arrival to block forged requests.
Tenant-Isolated RAG Vector Retrieval
The user query is converted into a vector and searched strictly against the requesting organization's private vector collection. Cross-tenant retrieval is architecturally impossible.
Enterprise LLM Inference (Zero Model Training)
A structured prompt is dispatched to Google Gemini Enterprise Agent Platform API. Under Google Cloud Service Terms (Section 17), customer prompts and responses are never used to train or fine-tune foundation AI models; processing is kept in transient memory without disk persistence.
Response Delivery & Asynchronous Audit Record
The AI response is returned to the customer, while the conversation exchange is asynchronously saved to the cloud database in Taiwan for agent review or human-in-the-loop takeover.
8. Data Lifecycle & Deletion Mechanisms
Orange AI provides complete data lifecycle management, ensuring data is appropriately protected at every stage and can be securely purged when no longer needed.
Data Creation
Uploaded knowledge documents immediately enter an isolated parsing pipeline, accessible only to authorized users within the organization. Temporary files are automatically cleaned up with a 1-day TTL.
Data Usage
During RAG queries, vector search is strictly scoped to the requesting organization. AI inference results are never used as training data.
Orange AI uses Google Gemini Enterprise Agent Platform Embedding (formerly Vertex AI, rebranded April 2026) to convert your knowledge documents into vectors. Per Google Cloud's terms of service, customer data sent via the Gemini Enterprise Agent Platform API is not used by Google to train its foundation models. Your enterprise knowledge remains under your full data sovereignty.
Data Retention
Conversation logs and audit records are retained per regulatory requirements. Exported audit reports (CSV etc.) are configured with automatic TTL expiry in encrypted object storage.
Organization Deletion & Data Erasure
When an organization account is terminated, the system automatically recursively deletes all associated database records, storage files, and vector data, leaving no residual data. This operation is executed by a Super Admin in the Admin Console with multiple confirmation steps to prevent accidental deletion.
9. Frontend Key Protection
Orange AI strictly adheres to frontend/backend separation. Sensitive API keys and business logic never appear in frontend code.
verified_user Frontend (Flutter Web / Web Widget) Design Principles
- check_circleContains no AI API keys (Gemini, OpenAI, etc.)
- check_circleContains no third-party OAuth Client Secrets
- check_circleNever calls AI inference APIs directly â all AI requests are proxied through backend microservices
- check_circleClient connectivity configuration is protected by App Attestation and server-side access control rules
functions Backend Microservice Security Management
- check_circleAll sensitive keys stored in Google Cloud Secret Manager (encrypted environment variables)
- check_circleIsolated execution environment â each request runs in an independent sandbox
- check_circleThird-party OAuth tokens are encrypted in dedicated databases; the frontend cannot read them directly
10. Rate Limiting & Abuse Prevention
To prevent malicious abuse or accidental API flood attacks, Orange AI deploys rate limiting mechanisms at multiple layers.
API Request Rate Limiting
Query endpoints enforce rate limits per session. Requests exceeding the threshold are automatically blocked, preventing any single user or automated script from exhausting shared compute resources.
File Upload Protection
File uploads are subject to strict size and format restrictions enforced at multiple layers. The parsing pipeline includes memory exhaustion protection to ensure service stability is not affected by large or malformed files.
Knowledge Compilation Resource Controls
Organization-level knowledge compilation tasks are subject to concurrency limits, preventing any single organization from monopolizing compute resources. Distributed locking ensures cross-tenant tasks do not interfere with each other.
Security Event Monitoring
Abnormal login attempts, repeated request failures, and suspicious operations are automatically recorded by the system for subsequent audit review and security alerting.
11. Audit Logs & Visual Operation Inspection
Orange AI delivers comprehensive operational auditing built specifically for enterprise governance and security compliance. In addition to immutable backend event logging, the Admin Console features an interactive, visual Audit Log Viewer and secure report export capabilities.
In-Console Audit Log Viewer & Visual Inspection
Tenant Administrators can access the dedicated Audit Logs module in the Admin Console to inspect, trace, and filter all critical events across their organization:
view_timeline Visual Interactive Timeline
Events are displayed in an intuitive chronological card flow, annotated with millisecond-precise timestamps, operator names, account emails, assigned roles, and action categories.
difference Field-Level Difference Inspection (Diff Card)
Integrated diff cards highlight exactly what changed between previous and new parameter values (e.g., prompt updates, model switches, RBAC adjustments), ensuring total change transparency.
date_range Dynamic Time Range Filters
Built-in time range presets (1 day, 3 days, 7 days, 14 days, 30 days) empower security teams to rapidly focus on specific operational windows and investigate anomalies.
file_download Cryptographically Signed URL CSV Export
Enables one-click CSV report exports. Backend microservices package the logs server-side and issue a time-limited (15-min TTL) secure signed download URL directly from storage, preventing unauthorized access.
Full Coverage Across 8 Operational Domains
The audit engine automatically logs actions across: 1. Organization Settings (plan updates, maintenance mode) ¡ 2. Agent Persona & Parameters (prompt edits, model selection, temperature, rate limit thresholds) ¡ 3. Knowledge Base & Wiki (document additions, deletions, manual compilation triggers) ¡ 4. Tool Integrations (enabling/disabling Google Calendar, Docs, Sheets, etc.) ¡ 5. Channel Webhooks (LINE, Meta, WhatsApp bindings/unbinds) ¡ 6. Membership & RBAC (invitations, role upgrades/downgrades, login/logout sessions) ¡ 7. Human Takeover Collaboration (dialog assignments, agent badges, manual responses) ¡ 8. Billing & Quota (add-on purchases and quota adjustments).
Server-Side Tamper Proof
Audit records are written strictly by trusted backend microservices. Server-side access control rules disallow frontend or user modification/deletion.
Strict RBAC Isolation
Only Tenant Administrators can inspect or export audit logs. Support staff and knowledge managers cannot access logs, preventing audit trail tampering.
Aggregated Monitoring & Alerts
System-level log volumes and failure events are regularly aggregated for dashboard metrics and automatic security alert notifications.
12. AI Prompt Security
To address security risks specific to Large Language Models (LLMs), Orange AI deploys dedicated protection mechanisms at the AI inference layer.
Prompt Injection Protection
System prompts and user input are strictly isolated. Enterprise-configured Agent rules are injected as Hard Rules that users cannot override through conversation content.
Multimodal Content Protection
Uploaded images and documents undergo format validation and binary explosion protection before entering the knowledge base, preventing malicious content from affecting the system through file uploads.
AI Loop Protection
Agent Tool Calling has a maximum iteration limit to prevent AI from entering an infinite loop when search results are insufficient, ensuring service stability.
Knowledge Base Boundary Isolation
RAG search vector space is strictly limited to the querying organization's scope. AI cannot "learn" from or access other organizations' knowledge base content.
13. Large Language Model Provider Policy
Orange AI uses Google's Gemini Enterprise Agent Platform (formerly Vertex AI, rebranded April 2026) as the core service for LLM inference and vector embeddings. The following outlines Google's explicit commitments and relevant terms regarding enterprise customer data.
verified_user Google Cloud Service Terms â Section 17: Training Restriction
Per Google Cloud's terms of service, Google will not use customer data â including prompts, outputs, and derived data â to train or fine-tune any AI/ML models without prior customer authorization. This commitment applies to all managed models on the Gemini Enterprise Agent Platform, including both GA and pre-GA models.
The terms further guarantee: your data will not be used to train models for any other Google customer.
How Your Data Is Handled
Default In-Memory Caching (No Disk Write)
Inference inputs and outputs are cached in-memory only by default â never written to disk. Data is isolated at the project level with a 24-hour TTL for automatic expiry, used solely to improve response latency.
Limited Retention for Safety & Compliance
In certain scenarios, Google retains request data for a limited period per its Cloud Data Processing Addendum (DPA), solely for security monitoring, abuse detection, and legal compliance â never for model improvement.
Zero Data Retention Option
Eligible enterprise customers can apply for Zero Data Retention terms by signing a Data Processing Addendum amendment with Google Cloud, providing an additional layer of data protection.
Important: Paid API vs. Free Tier
The training restrictions and data protection commitments above apply exclusively to the paid Gemini Enterprise Agent Platform API. The free Google AI Studio tier is not covered by these protections. Orange AI exclusively uses the paid enterprise-grade API to ensure your data is fully protected.
References: Gemini Enterprise Agent Platform Zero Data Retention ¡ How Gemini for Google Cloud uses your data ¡ Google Cloud Service Terms
14. Third-Party Service Security
Orange AI integrates multiple third-party services to provide complete functionality. We apply the least-privilege principle to every integration and ensure their security at the architectural level.
| Service | Purpose | Security Measures |
|---|---|---|
| Google Gemini AI + Gemini Enterprise Agent Platform |
AI inference, vector embedding | API key stored in Secret Manager only; called via backend proxy. Google commits that Gemini Enterprise Agent Platform API data is not used for model training. |
| Meta (Facebook/IG) | Social messaging channel webhooks | Webhook verification token encrypted; source IP validation |
| Google Calendar/Docs/Sheets | Tool integrations | OAuth 2.0 tokens encrypted at rest; minimum required scopes per org |
| Cloud Push Notification Service (FCM) | Agent push notifications | Device tokens bound to specific users; regular rotation mechanism |
| Apple Sign-In | iOS social login | Race condition protection; strict credential configuration control |
15. OAuth & Credential Management
All third-party OAuth authorization flows are completed server-side, ensuring Client Secrets are never exposed to the frontend.
Multi-Tenant OAuth Isolation
Each organization's OAuth tokens are stored in that organization's dedicated database partitions. Backend microservices from other organizations cannot read cross-tenant tokens. Frontend paths for directly reading OAuth tokens are completely blocked by server-side access control rules.
Google Workspace integrations (Calendar, Docs, Sheets) comply with Google API's Limited Use policy, requesting only the minimum OAuth scopes required to deliver the functionality. When users revoke authorization, related features are immediately disabled.
16. Regulatory Compliance & Enterprise NDA Safeguards
Data Protection & Privacy
-
gavel
Taiwan PDPA (Personal Data Protection Act)
Primary databases and microservices operate locally in Taiwan, strictly complying with domestic data residency and personal data collection and processing laws. -
gavel
GDPR Compatible
Underlying infrastructure is GDPR-compliant by design, fully supporting tenant data portability and the Right to Erasure. -
gavel
EU Cyber Resilience Act (CRA) Compliance
In compliance with Regulation (EU) 2024/2847, enforcing end-to-end software supply chain security, daily automated dependency vulnerability scans, and standardized machine-readable SBOM generation.
Infrastructure Certifications
- verified Google Cloud ISO 27001 Certified
- verified SOC 1 / SOC 2 / SOC 3 Compliant
- verified Google API Limited Use Policy Compliant
- verified Apple App Store Review Guidelines Compliant
EU CRA Software Bill of Materials (SBOM) & Supply Chain Resilience
In accordance with the mandatory requirements of the EU Cyber Resilience Act (CRA - Regulation (EU) 2024/2847) for digital products and enterprise software lifecycle security, Orange AI implements a 100% cloud-native automated supply chain defense pipeline to guarantee our platform operates free of unmonitored third-party component risks:
account_tree Dual-Standard Machine-Readable SBOM
Automatically generates machine-readable JSON SBOMs adhering to both ENISA-recommended CycloneDX 1.6 and international standard SPDX 2.3 (ISO/IEC 5962), covering frontend, backend, SDK, and mobile native dependencies.
published_with_changes Daily Continuous Vulnerability Monitoring
Powered by GitHub Actions scheduled pipelines, running daily automated audits against global vulnerability databases (NVD, GitHub Security Advisories, and Google OSV), with structured SARIF reports ingested into repository security monitoring.
crisis_alert 24-Hour Critical Incident Response SOP
Pursuant to CRA Article 14, unmitigated CRITICAL or actively exploited vulnerabilities trigger automated CI build breaks and real-time alerts, initiating statutory reporting and rapid remediation procedures within 24 hours.
inventory 90-Day Audit Retention & Request Workflow
Historical SBOMs and vulnerability audit logs are securely encrypted and retained for 90 days. Enterprise clients and regulators can request the latest verified SBOM anytime via account managers or at service@orangeai.tw.
Enterprise Non-Disclosure Agreement (NDA) Signing & Template Support
Orange AI deeply recognizes enterprise concerns regarding trade secrets, proprietary knowledge, business workflows, and customer communication data. We provide robust contractual confidentiality protections for all prospective and active clients:
assignment_turned_in Standard Mutual NDA Template Available
We provide a comprehensive, attorney-drafted standard Mutual Non-Disclosure Agreement (MNDA) template that clearly governs confidential information definitions, non-disclosure covenants, standard exclusions, and survival terms.
description Preview & Print Standard MNDA Templateedit_document Client Custom Legal Templates Supported
If your organization requires its own vendor or partner confidentiality terms, our legal team can review and execute your proprietary NDA template to streamline procurement and technical evaluation.
mark_email_read How to Request and Execute an NDA
Whether during early Proof of Concept (PoC), technical evaluation, or commercial deployment, you can contact your dedicated account manager or email service@orangeai.tw with the subject "Request / Sign Enterprise Non-Disclosure Agreement (NDA)". We will provide the agreement and facilitate digital signatures (DocuSign / Adobe Sign) or physical execution within 1 to 2 business days.
17. Security Vulnerability Disclosure
Orange AI values all reports from users and security researchers. If you discover a potential security vulnerability, please contact our security team through the following channels. We commit to an initial response within 72 hours of notification and will remediate confirmed vulnerabilities within a reasonable timeframe.
Security Contact
Responsible Disclosure Policy: We respect the contributions of security researchers. Please do not publicly disclose vulnerability details before a fix is in place. We do not take legal action against good-faith security research and welcome collaboration with the research community.
Š 2026 Orange AI ¡ Security Whitepaper v1.1 ¡ Contact service@orangeai.tw
| |