AI Security
Securing AI Applications: Prompt Injection, Enforcement and Evidence
Course Overview
AI-enabled applications introduce security challenges when untrusted content is processed alongside legitimate user requests and application instructions.
Prompt injection attempts may appear in documents, retrieved records, external content, or tool responses.
Detecting suspicious content is an important security capability, but detection alone does not establish that an application prevented an unauthorized action.
This proposed SecureStack™ curriculum examines the distinction between identifying prompt injection attempts, enforcing application security boundaries, and collecting evidence that demonstrates what actually occurred.
The Central Security Question
Was the malicious instruction detected, and did the application enforce the intended boundary?
Proposed Learning Objectives
Participants should be able to:
Distinguish direct prompt injection from indirect prompt injection.
Identify untrusted content entering AI-enabled application workflows.
Explain the difference between detection and enforcement.
Evaluate whether suspicious content was blocked before reaching a model.
Interpret evidence showing application and model execution outcomes.
Assess how downstream tool authorization provides an additional security boundary.
Proposed Curriculum Structure
Module 1 — AI Application Trust Boundaries
- Application instructions
- User prompts
- Retrieved content
- External documents
- Tool responses
Module 2 — Prompt Injection Threats
- Direct prompt injection
- Indirect prompt injection
- Embedded malicious instructions
- Instruction versus data boundaries
Module 3 — Detection Controls
- Prompt injection detection
- Azure AI Content Safety Prompt Shields
- Detection results
- Security signal interpretation
Module 4 — Application Enforcement
- Application-level blocking
- Pre-model execution controls
- Security decisions
- Model invocation evidence
Module 5 — Tool-Output Security
- AI agents consuming tool responses
- Post-tool guardrails
- Suspicious retrieved content
- Preventing unauthorized instruction following
Module 6 — Evidence and Validation
- Detection evidence
- Enforcement evidence
- Model invocation status
- Tool execution results
- Distinguishing observed outcomes from assumptions
Proposed Technical Foundations
These proposed technical foundations do not represent that all proposed course exercises have been developed or validated.
Offering Status
- Portfolio status
- Proposed — curriculum development
- Category
- AI Application Security
- Delivery format
- VILT and ILT engagements
- Registration
- Not open
This proposed curriculum is under development. Modules, exercises, and technical demonstrations described on this page are proposed and have not been validated.
Proposed Technical Foundations
Proposed Offering
Securing AI Applications is a proposed SecureStack™ offering. It is not open for registration, and no dates, pricing, or delivery schedule have been set.
Explore the SecureStack™ Course Catalog →