AIS103 · Securing AI Systems and LLM Applications
About Course
Protect AI systems and LLM applications with a defense-in-depth approach. You threat-model an LLM app using the OWASP Top 10 for LLM Applications and MITRE ATLAS, apply layered controls and verify them with automated evaluations.
Who it is for
IT and security professionals, developers and product owners responsible for deploying chatbots, assistants or AI agents.
Prerequisites
Complete AIS101; basic familiarity with web applications and reading YAML or JSON is helpful.
What you will be able to do
- Threat-model an LLM application using the OWASP Top 10 for LLM Applications (2025) and MITRE ATLAS
- Design layered controls for input, retrieval, model, output and tool use
- Apply least privilege and human approval to AI agents and integrations
- Verify defenses with an automated evaluation suite and monitor AI applications in production
Lessons
- Lesson 1 · The attack surface of AI and LLM applications
- Lesson 2 · Layered defenses for LLM applications
- Lesson 3 · Securing agents, tools and the AI supply chain
- Lesson 4 · Lab and assignment
Lab: Threat-model and harden a support chatbot
You draw the architecture of a customer-support chatbot that answers from a knowledge base and can look up order status. You build a threat model, write a hardened system prompt and output checks, and run a promptfoo evaluation suite with its built-in safety and data-leak test categories against your local model before and after your mitigations.
Free tools: Ollama with a small open model (free, runs locally), promptfoo (open source), VS Code and draw.io
Assessed assignment
Write a security design review for an LLM-powered assistant at a small online retailer that answers customer questions, searches internal policy documents and can create return labels. Include an architecture diagram with trust boundaries, a threat model covering at least six OWASP LLM risks, the controls you would apply at each layer, the permissions given to each tool, and your evaluation results before and after mitigations. End with a monitoring plan.
| Criterion | Weight | What excellent work looks like |
|---|---|---|
| Architecture and trust boundaries | 20% | The diagram shows all components, data stores, tools and boundaries where untrusted content enters. |
| Threat model | 30% | At least six relevant OWASP LLM risks are analyzed with realistic scenarios and impact. |
| Layered controls and least privilege | 30% | Controls are specific, cover every layer and limit tool permissions with approvals where needed. |
| Evaluation and monitoring | 20% | Before-and-after results are documented and the monitoring plan defines signals, thresholds and owners. |
Portfolio piece
A security design review with threat model, controls and evaluation results for an LLM application.
Next step
Continue with AIS104 · Responsible AI: Bias, Transparency and Human Oversight · 6 hours · $42.00
Course Content
AIS103 · Securing AI Systems and LLM Applications (6 h)
-
Lesson 1 · The attack surface of AI and LLM applications
-
Lesson 2 · Layered defenses for LLM applications
-
Lesson 3 · Securing agents, tools and the AI supply chain
-
Lesson 4 · Lab and assignment
-
AIS103 Assessment · Securing AI Systems and LLM Applications