Unit 01 · 1 h 28 min
How LLMs work, for defenders
Tokens, context windows, temperature and system prompts, seen from the side that has to secure them.
AI and LLM Security
Secure the AI assistants, agents and models your company ships, with a live LLM app to test, trace and fix.
- Time
- 16 h 54 min
- Lessons
- 61
- Machine
- Ubuntu 24.04 LTS
What you earn
RootCertificateAI and LLM Security, certificate of completion IAPPAIGPArtificial Intelligence Governance Professional
Prepares you for Artificial Intelligence Governance Professional. Root is not affiliated with these issuers.
You will be able to
- Map an LLM app's prompts, retrieval, tools and integrations
- Detect leaks and injected content with canaries, quarantine and tests
- Put agents, tools and MCP servers behind least privilege and approval
- Find shadow AI use and set rules staff can follow
- Close exposed model servers, vector stores and shared links
Units and lessons
Unit 02 · 1 h 49 min
Anatomy of an LLM app
Take a real looking AI assistant apart: prompt builder, retrieval, tools, MCP servers and logs.
Unit 03 · 1 h 32 min
AI at work: use policy and governance
Set the rules for how staff use AI tools, find the use that is already happening, and keep an inventory of every AI system.
- 1Case file: what staff paste into AI tools18 min · machine
- 2Find AI use you did not approve16 min · machine
- 3Compare AI plans before you approve them14 min · virtual lab
- 4Case file: a payment request with a face16 min · virtual lab
- 5Keep an inventory of every AI system14 min · virtual lab
- 6Review AI vendors before you renew14 min · virtual lab
Unit 04 · 1 h 28 min
Prompt injection: detect and contain
Run your app's own guardrail tests, catch leaks with a canary token and trace probing in the request log.
Unit 05 · 1 h 29 min
Untrusted content: documents, pages and email
Find test fixtures in retrieved documents, web pages and email, and quarantine them before the model reads them.
Unit 06 · 1 h 23 min
Sensitive data leakage
Find secrets in configs and prompts, card numbers in logs, and prove the output guard redacts them.
Unit 07 · 1 h 27 min
Excessive agency: tools, approval and MCP
Cut tools down to least privilege, put money and email behind approval, and audit what agents actually did.
Unit 08 · 1 h 20 min
AI coding assistants and agents on developer machines
Lock down the agents developers run: permission modes, secrets they can read, MCP servers, skills, extensions and the signs of abuse.
Unit 09 · 1 h 13 min
Model and AI supply chain
Check downloaded models before anything loads them: formats, pickle contents, hashes and model cards.
Unit 10 · 58 min
Exposed AI infrastructure
Find self hosted model servers, vector databases and shared links that answer anyone, and close them.
Unit 11 · 1 h 19 min
Monitoring and evaluating AI apps
Log the right fields, catch runaway usage, rerun guardrail tests on every change and triage alerts.
Unit 12 · 1 h 28 min
Capstone: audit an AI app against the OWASP LLM Top 10
Map the OWASP Top 10 for LLM applications, audit Atlas, fix every finding and prove it with a clean rerun.
Take apart a working AI support assistant, govern how staff and developers use AI, and close the infrastructure around it, with every unit built around real incidents. Finish by auditing two apps against the OWASP Top 10 for LLM applications.