AI Security Foundations

What you will be able to do

• You will be able to map the attack surface of an AI system across its data, model, and deployment stages.
• You will be able to explain the core adversarial machine learning attacks: evasion, poisoning, model extraction, and inference attacks.
• You will be able to assess a large language model application against the OWASP Top 10 for Large Language Model Applications.
• You will be able to identify prompt injection and insecure output handling in a real LLM feature and recommend concrete mitigations.
• You will be able to describe an AI attack as a chain of tactics and techniques using the MITRE ATLAS framework.
• You will be able to recognize how classic risks like supply chain weakness and excessive agency show up in AI systems.
• You will be able to translate a mapped threat into prioritized, defensible security controls.
• You will be able to produce an AI security assessment that a stakeholder can act on.

What is inside

4 modules, 8 lessons. Each module ends in a graded quiz and most carry an assignment.

1. 01

   The AI Attack Surface

   AI and machine learning systems fail in ways classic application security never had to model, because their behavior is learned from data rather than written in code. This module walks the full pipeline, from training data to the deployed model to the applications built on top, and shows where each stage can be attacked and why. You finish with a clear map of the AI attack surface, the attacker goals behind it, and the shared vocabulary this field uses.

   2 lessons · 5 quiz questions

2. 02

   Adversarial Machine Learning

   Adversarial machine learning is the study of how models can be fooled, corrupted, copied, or made to leak, and it is core threat knowledge for any AI security analyst. This module works through the four attack families every analyst must recognize: evasion and poisoning, which target how a model predicts and how it is trained, and model extraction and inference attacks, which steal the model itself or the private data behind it. For each family you will see how the attack actually works, what it costs an attacker, and where defenses realistically fit.

   2 lessons · 5 quiz questions · assignment

3. 03

   OWASP Top 10 for LLM Applications

   Large language model applications carry their own ranked list of risks, published by OWASP (the Open Worldwide Application Security Project). This module works through the Top 10 in two lessons. The first covers the risks closest to the prompt: prompt injection, insecure output handling, and sensitive information disclosure. The second covers the rest, from supply chain and excessive agency to overreliance and model theft. Throughout, you will practice spotting each risk in a live LLM feature and reasoning clearly about the controls that reduce it.

   2 lessons · 5 quiz questions

4. 04

   MITRE ATLAS and the Analyst Workflow

   MITRE ATLAS gives AI attacks a shared language of tactics and techniques, the same way ATT&CK did for enterprise defenders. This module teaches you to read the ATLAS matrix, map a real attack and system onto its tactics and techniques, turn that map into a short list of prioritized defenses, and write it up so a stakeholder can act. You finish by running the full analyst workflow end to end against a sample AI system.

   2 lessons · 5 quiz questions · assignment