Security & Infrastructure Whitepaper

Last updated: November 20, 2026

AI Course Generation Platform · Version 1.0 · 2026

Executive Summary

This document provides a comprehensive overview of the security architecture, data handling practices, infrastructure controls, and privacy safeguards implemented within the AI Course Generation Platform. It is intended for prospective customers conducting technical due diligence and security evaluations.

The AI Course Generation Platform enables organizations to automatically generate structured e-learning content — including course topics, slides, quizzes, audio summaries, and supporting media — through a multi-stage pipeline powered by large language models (LLMs), licensed stock-image search, vector embeddings, and text-to-audio synthesis. Every stage of this pipeline has been designed with security-first principles to ensure the confidentiality, integrity, and availability of customer data.

KEY

All AI processing is scoped to individual course contexts. Customer documents and generated content are never shared across tenants, exposed to internet-accessible endpoints, or used for model training purposes.

This whitepaper addresses the following security domains: authentication and access control, data isolation and multi-tenancy, storage and encryption, AI pipeline security, document handling and embeddings, network security, and compliance readiness.

Platform Architecture Overview

The AI Course Generation Platform consists of two primary content generation pathways — AI-Generated Content and User Document-Driven Content — both converging on a common output pipeline. Understanding this architecture is foundational to evaluating the platform's security posture.

Content Generation Pathways

Pathway 1: AI-Generated Content

In this pathway, all source material is derived from pre-configured, platform-managed databases. The flow proceeds as follows:

  • Course topics and descriptions are retrieved from a secured database, accessible only via authenticated, defined application paths.
  • Approved large language models generate course names, topic outlines, slides, quizzes, and summaries based on administrator-defined industry and topic selections. The platform supports Anthropic Claude and Mistral models through centrally managed, task-specific model configurations.
  • Licensed stock-image services produce course thumbnails and slide visuals.
  • Text-to-audio services synthesize spoken summaries.
  • All database queries for industry options and slide selections are secured and preset — end users cannot inject arbitrary queries.

The application manages model selection and provider interaction through the backend. This enables the platform to apply consistent controls for authenticated access, prompt construction, usage management, monitoring, and operational resilience across supported AI providers.

Pathway 2: User Document-Driven Content

In this pathway, customers supply their own proprietary documents (e.g., PDFs, internal training materials) to ground AI-generated content. This pathway carries elevated data handling responsibilities:

  • Documents are uploaded by authenticated users and stored in isolated S3 buckets scoped to the user level.
  • Alternatively, documents can be selected from course-level asset libraries, further restricting access scope.
  • Embeddings are generated from document content and stored in a vector database, isolated at the course level.
  • LLM calls reference retrieved document segments (paragraphs, content blocks) to produce contextually grounded course content.
  • No document content is accessible outside the user/course scope boundary.

User-supplied documents are treated as source material for the course-generation workflow. Retrieved document content is used as contextual reference material and remains distinct from the platform's application instructions and generation rules.

Scope Isolation Model

The platform enforces a two-level scoping model that governs data access throughout the pipeline:

Scope LevelWhat It ControlsAccess Boundary
Course LevelEmbeddings, vector DB records, asset selections, slide DB queriesIsolated per course; no cross-course data leakage
User LevelUploaded documents stored in S3Accessible only to the authenticated user who uploaded them

This dual-scope model ensures that even within the same organization or tenant, course content and user documents are segregated at the data layer.

Authentication and Access Control

Authenticated Access Requirement

All platform functionality — including document uploads, AI pipeline execution, and content retrieval — requires authenticated access. There are no publicly accessible endpoints that expose customer data or AI-generated content. Only authenticated users can access the documents.

The platform integrates with the organization's LMS to establish authenticated user sessions. Protected capabilities are available only through authenticated application flows, and backend services validate identity information before allowing access to course creation, asset management, generation, and retrieval functions.

Database Access Controls

Platform databases (topic databases, industry options, slide selections) are configured with the following security controls:

  • Access is restricted to authenticated, application-defined code paths only.
  • Direct internet access to database resources is not permitted at any layer.
  • Database credentials and connection strings are managed at the application infrastructure level and are not exposed to end-user sessions.
  • Queries are pre-defined and parameterized — ad hoc or user-supplied database queries are not supported.

Document and Asset Access

User-uploaded documents stored in S3 are subject to strict access controls:

  • S3 bucket policies are scoped to individual users, preventing cross-user document access.
  • Course-level asset selections are governed by course-scoped permissions, ensuring content reuse is limited to authorized courses.
  • Access to documents and course assets is managed through authenticated application workflows and authorized course context. The platform is designed to ensure that protected content is accessed only by users with appropriate permissions.

Data Isolation and Multi-Tenancy

Course-Level Data Isolation

The platform implements application-layer data isolation at the course scope. Each course's embeddings and vector database records are stored with course-specific identifiers, preventing any retrieval mechanism from returning data belonging to a different course. This applies to:

  • Vector database queries used to retrieve document segments for LLM context injection
  • Slide selection database queries scoped to per-course configurations
  • Image assets associated with specific course generation runs

User-Level Document Isolation

Documents uploaded by users are stored in user-scoped S3 paths. The isolation guarantees are:

  • No user can access documents uploaded by another user, even within the same organization.
  • Document content is not indexed in any shared or global search index.
  • Documents uploaded during a session are not accessible from prior or future sessions of other users.

AI Pipeline Data Boundaries

A critical security property of the platform is that AI pipeline processing (LLM inference, embedding generation, image synthesis) is scoped to the active course context. Specifically:

  • LLM prompts contain only course-relevant content: the course name, industry, description, selected titles, and retrieved document segments from the current course's vector store.
  • There is no mechanism by which content from one user's uploaded documents could appear in another user's LLM context window.
  • Stock-image search requests carry only a short search phrase, derived from the slide title and image hint, or the course name, together with display filters such as result count and orientation. No document content, document excerpt, filename, or user identifier is transmitted to an image provider.

AI processing is purpose-limited and course-aware. Each generation request uses only the context required for the specific task, such as an outline, a selected slide, or approved source material associated with the active course. This approach supports relevant course generation while reducing unnecessary data exposure.

Storage Security and Encryption

S3 Document Storage

User-uploaded documents are stored in Amazon S3 with the following security configuration:

  • Buckets are configured with server-side encryption (SSE) enabled at rest.
  • Bucket policies enforce access only through authenticated application service principals.
  • Public access is blocked at both the bucket and account level.
  • S3 Transfer Acceleration and HTTPS-enforced bucket policies ensure encryption in transit for all upload and retrieval operations.

Vector Database Storage

Document embeddings stored in the vector database represent a mathematical transformation of source content and are scoped at the course level. Security controls include:

  • Vector records are tagged with course-level identifiers to enforce retrieval isolation.
  • The vector database is not internet-accessible; all queries are executed from within the application's private network boundary.
  • Each stored record comprises the numerical embedding, the source text segment that produced it, and scope metadata (course identifier, source filename, file identifier, segment index). The text segment is retained because retrieval must return readable text to ground generation — an embedding is not reversible into text. Because the vector database is self-hosted on the platform's private network, this retained text remains inside the platform's controlled infrastructure at all times and is never held by a third-party vector service.

Relational Database Security

Preset databases (initial topics, industry options, slide selection parameters) are:

  • Accessible only via authenticated, application-defined connection paths.
  • Not directly accessible from the internet or via ad hoc query tools.
  • Secured with encrypted connections and role-based access controls aligned to least-privilege principles.

AI Pipeline Security

LLM Integration (Mistral)

The AI Course Platform uses a provider-agnostic large language model architecture. The platform currently supports Anthropic Claude and Mistral models through a centralized model registry, with optional OpenAI support where enabled by deployment configuration.

Model selection is task-based. Different tasks, including slide generation, outline generation, quiz generation, course recommendations, brand-kit generation, and audio suggestions, may use different models according to configured priority chains. The platform can fall back to another approved model or provider when a configured provider is unavailable, subject to the applicable retry, rate-limit, and authorization rules.

The platform supports Claude Opus, Claude Sonnet, Claude Haiku, and Mistral model configurations. Provider and model selection is governed through controlled deployment configuration, allowing the platform to select the most appropriate approved model for each workload while preserving consistent application-level security controls.

The provider abstraction allows the platform to evolve its model portfolio without requiring changes to the surrounding authentication, authorization, data-protection, monitoring, and audit architecture. Provider selection remains subject to approved configuration, contractual requirements, data-protection obligations, and the security controls described in this whitepaper.

Prompt Construction

  • Prompts are constructed programmatically using only validated, sanitized inputs from the platform's own database or user session context.
  • User-supplied free-text input is not directly interpolated into LLM prompts without sanitization and contextual scoping.
  • Uploaded document content is accessed only via the vector retrieval layer, which returns pre-chunked, course-scoped segments — reducing the risk of prompt injection via malicious document content.

Data Minimization in LLM Calls

Each AI request is assembled for a defined generation task and includes only the context needed to produce the requested output. Course metadata, selected learning objectives, approved source material, and relevant outline or slide context may be used where appropriate. The platform does not provide models with unrestricted access to unrelated course data, application credentials, or customer content outside the active workflow.

Pipeline StageData Sent to LLM
Course NamesCourse description (preset or user-defined)
Ten TopicsDescription, course name, industry
OutlineCourse name, selected paragraphs from uploaded documents
Slides + QuizCourse name, title, description, retrieved document paragraphs
SummarySlide content from current course only

No Model Training on Customer Data

Customer-uploaded documents and AI-generated course content are not used for retraining, fine-tuning, or improving any AI model. LLM API calls are stateless inference requests; no persistent memory or learning occurs from customer interactions.

Image Generation Security

Course thumbnails, slide photography, and icons are sourced from licensed stock libraries by keyword search. The platform does not operate, call, or embed any image-generation model. No image prompt is constructed, and no image model is queried, trained, or fine-tuned.

  • Primary provider: Freepik/Magnific, queried for photo content types with an optional orientation filter.
  • No document content, document excerpt, filename, user identifier, course identifier, or session token is transmitted to either provider. The full extent of what a stock provider observes is a short search phrase and the platform's API credential.

Text-to-Audio Security

The text-to-audio pipeline converts LLM-generated summary text to audio. The following controls apply:

  • Input to the audio synthesis service is limited to the AI-generated summary text — not the original uploaded documents.
  • Audio output files are stored in course-scoped storage and are not publicly accessible.
  • The service endpoint is accessed from within the application's private network perimeter.
  • Primary provider: Murf

Network Security and Infrastructure Controls

Implementation-Level Isolation

The architecture designates an Implementation Level for all AI processing services (LLM inference, embedding generation, vector DB queries, image generation, audio synthesis). Components at this level:

  • Cannot be accessed from the internet directly.
  • Are not reachable via direct URLs, SCORM packages, LMS integrations, or any external sharing mechanism.
  • Communicate exclusively through application-internal service calls authenticated at the infrastructure layer.

Storage-Level Controls

The Storage Level (encompassing S3 document storage, asset management, and document-to-asset selection) enforces:

  • Authentication for all read and write operations.
  • No anonymous or unauthenticated access paths.
  • Encryption in transit enforced via HTTPS/TLS for all API calls.

Defense-in-Depth Architecture

The platform implements a layered defense strategy. The following table summarizes the security controls at each architectural layer:

LayerComponentSecurity Control
PresentationUser-facing UI/APIAuthenticated sessions, HTTPS-only
ApplicationPipeline orchestrationInput validation, scope enforcement, parameterized DB queries
AI ServicesLLM, Image, Audio APIsPrivate network access, no direct internet exposure, data minimization
DataS3, Vector DB, Relational DBEncryption at rest and in transit, course/user scoping, access policies
InfrastructureCloud hosting environmentIAM roles, least-privilege service principals, no public DB endpoints

Customer Document Handling

Upload and Ingestion

When a customer uploads a proprietary document, the following sequence of security-relevant operations occurs:

  • The document is transmitted over HTTPS to the platform's authenticated upload endpoint.
  • The document is stored in a user-scoped S3 path with server-side encryption applied immediately upon storage.
  • Alternatively, documents may be promoted to course-level asset libraries by authorized users, at which point access scope is expanded to the course level (not beyond).

Embedding Generation and Vector Storage

Document embeddings are generated through the following controlled process:

  • Documents are chunked into segments optimized for semantic retrieval.
  • Each chunk is encoded into a multi-dimensional vector representation by an embedding model.
  • Vectors are stored in the vector database tagged with course-level scope identifiers.
  • Raw document text is not stored in the vector database — only numerical embeddings.
  • Retrieval operations are scoped to the active course's namespace, preventing cross-contamination with other courses' document content.

Data Retrieval for Content Generation

During course content generation, the following retrieval controls are enforced:

  • Vector similarity queries are executed only within the course-scoped namespace.
  • Retrieved segments are passed to LLM prompts as grounding context — they are not stored, logged, or reused across sessions.
  • The number of retrieved segments per LLM call is bounded to prevent excessive data exposure.

Privacy, Compliance, and Data Governance

Data Minimization Principles

The platform is designed in accordance with data minimization principles. At each stage of the AI pipeline, only the data strictly required for that processing step is accessed, transmitted, and retained. Customer documents are not retained in intermediate processing caches beyond the active generation session.

Data Residency and Sharing

The platform's architecture enforces strict boundaries on data sharing and residency:

  • Customer-uploaded documents and generated course content are not shared with third parties for any purpose other than the specific AI API calls required for generation.
  • Third-party AI and media service calls — including approved Anthropic Claude and Mistral model configurations, image services, and text-to-audio services — receive only the information required for the relevant generation task.
  • Generated content is not shared outside the authenticated user and course scope via any technical mechanism (direct links, LMS publish, SCORM, etc.).

Audit and Logging

The platform supports comprehensive logging of security-relevant events to support audit requirements:

  • User authentication events (login, logout, session expiry)
  • Document upload and retrieval operations, tagged with user and course identifiers
  • AI pipeline execution events, including inputs and completion status (without storing raw LLM prompt content)
  • Database access events from application service principals

Compliance Framework Alignment

The security controls described in this whitepaper are designed to support compliance with common enterprise security frameworks. The following table summarizes alignment:

FrameworkRelevant DomainPlatform Alignment
SOC 2 Type II
(Certification Pending)
Security, Availability, ConfidentialityAccess controls, encryption, data isolation, audit logging
ISO 27001
(Certification Pending)
Information Security ManagementAsset management, access control, cryptography, supplier security
GDPR / CCPAData PrivacyData minimization, access scoping, no third-party sharing
NIST CSFIdentify, Protect, Detect, RespondLayered controls, authentication, monitoring, incident response readiness
INFO

Customers requiring formal compliance certifications (e.g., SOC 2 Type II report, penetration test results, or DPA documentation) should contact their account executive to request the relevant documentation under NDA.

Shared Responsibility Model

Security is a shared responsibility between the platform provider and the customer. The following table clarifies the division of responsibilities:

Responsibility DomainOwner
Platform infrastructure, encryption, AI pipeline controls, data isolationPlatform Provider
User credential management and access provisioning for your organizationCustomer
Content of uploaded documents and appropriateness for processingCustomer
Compliance obligations specific to your industry or jurisdictionCustomer
Security configuration of your LMS or SCORM-consuming systemCustomer
Network security of your organization's access to the platformCustomer

Conclusion

The AI Course Generation Platform has been architected to meet the security and data privacy expectations of enterprise customers handling sensitive internal training content. Key security properties include:

  • Complete isolation of customer data at the user and course scope levels
  • Authentication required for all data access operations — no public endpoints
  • Implementation-level AI processing services inaccessible from the internet
  • Data minimization in all AI API calls — sensitive document content never leaves the platform's controlled environment unnecessarily
  • Encryption at rest and in transit across all storage and communication layers
  • No use of customer data for AI model training or improvement purposes

We welcome the opportunity to discuss any aspects of this security architecture in greater detail. For technical deep-dives, vendor security questionnaires, or access to supporting compliance documentation, please contact your customer relationship rep.