Securing Retrieval-Augmented Generation and LLM Applications

IntermediateFree5 modules · 15 lessons

About this course

A working method for securing applications built with retrieval-augmented generation (RAG) and other large language model (LLM) architectures. You will map the RAG architecture stage by stage, enforce access control at the retrieval layer so it matches what a direct query against the underlying store would enforce, close data leakage paths that survive correct access control, secure the embedding pipeline and vector store as a write target in its own right, and build evaluation and monitoring that catches retrieval-specific failures generic application monitoring misses. Prompt injection payload technique is covered elsewhere and cross-referenced rather than repeated; this course focuses on the architecture, access control, data handling, and operational practice around retrieval-augmented systems.

What you'll learn

RAG Architecture and Why It Is a Distinct Risk

What Retrieval-Augmented Generation Actually Is
Why RAG Introduces a Distinct Risk Beyond Prompt Injection
The RAG Threat Model End to End

Access Control at the Retrieval Layer

Why Retrieval Must Enforce the Same Permissions as a Direct Query
The Common Mistake: Application-Level Checks the Retrieval Layer Bypasses
Designing Permission-Aware Retrieval

Data Leakage Through Retrieved Context

How Sensitive Content Ends Up in a Response to the Wrong User
Cross-Tenant and Cross-User Leakage in Shared Knowledge Stores
Output Handling for Generated Responses That Echo Retrieved Context

01RAG Architecture and Why It Is a Distinct Risk3 lessons

What Retrieval-Augmented Generation Actually Is
Why RAG Introduces a Distinct Risk Beyond Prompt Injection
The RAG Threat Model End to End

02Access Control at the Retrieval Layer3 lessons

Common questions

How much does Safeguard Academy cost?

Nothing. Every course, exam, and certificate on the catalog is free — including retakes. All you need is a free Safeguard account.

What are the prerequisites?

None. The flagship course, Safeguard Certified Practitioner, is a beginner-level course — basic familiarity with how software is built helps, but every exam question is answerable from the lessons themselves.

Embedding and Vector Store Security

Who Can Write to the Vector Index and Why That Matters
Stale and Malicious Embeddings and Their Persistence
Securing the Embedding Pipeline End to End

Evaluation and Monitoring for RAG Applications

Why Generic Web Application Monitoring Misses Retrieval-Specific Failures
Building a RAG-Specific Evaluation and Monitoring Program
A Worked Incident Walkthrough: Responding to a Suspected Retrieval Poisoning Event

Why Retrieval Must Enforce the Same Permissions as a Direct Query

The Common Mistake: Application-Level Checks the Retrieval Layer Bypasses

Designing Permission-Aware Retrieval

03Data Leakage Through Retrieved Context3 lessons

How Sensitive Content Ends Up in a Response to the Wrong User
Cross-Tenant and Cross-User Leakage in Shared Knowledge Stores
Output Handling for Generated Responses That Echo Retrieved Context

04Embedding and Vector Store Security3 lessons

Who Can Write to the Vector Index and Why That Matters
Stale and Malicious Embeddings and Their Persistence
Securing the Embedding Pipeline End to End

05Evaluation and Monitoring for RAG Applications3 lessons

Why Generic Web Application Monitoring Misses Retrieval-Specific Failures
Building a RAG-Specific Evaluation and Monitoring Program
A Worked Incident Walkthrough: Responding to a Suspected Retrieval Poisoning Event

What happens if I fail the exam?

You can retake it after a 24-hour cooldown, as many times as you need. Retakes are free, and each attempt draws a fresh random set of questions.

How long is a certificate valid?

The Safeguard Certified Practitioner credential is valid for 24 months from issue. The expiry date is printed on the certificate and shown live on its public verification page. Renewing means passing the current exam again.