MITRE ATLAS: Securing AI Systems Against Adversarial Attacks

“This course contains the use of artificial intelligence.”

MITRE ATLAS (Adversarial Threat Landscape for Artificial-Intelligence Systems) is a knowledge base of real-world adversary tactics, techniques, and case studies targeting machine learning and AI systems, modeled on MITRE ATT&CK and built from red-team observations. This course teaches you to use ATLAS to threat-model your own AI systems, not just to study attacks in the abstract.

You will walk the full ATLAS matrix tactic by tactic, from Reconnaissance and Resource Development through AI-specific tactics like ML Model Access and ML Attack Staging, all the way to Collection, Exfiltration, and Impact, and map a real AI attack onto it. Then you will get hands-on with the core adversarial ML attack classes aligned to the NIST AI 100-2 taxonomy: evasion (adversarial examples at inference), data poisoning and backdoors, model extraction and stealing, model inversion, and membership inference. A focused section covers how ATLAS now catalogs generative-AI and LLM threats, including prompt injection, jailbreaks, meta prompt extraction, and plugin and supply-chain compromise.

Crucially, every attack is paired with a defense. You will map ATLAS mitigations to techniques, apply adversarial training and model hardening, protect data and privacy, limit model access, and secure your ML supply chain. Finally, you will learn to run an ATLAS-based threat modeling workshop, prioritize risks with the ATLAS Navigator and case studies, build an AI red-teaming practice, and align your program with the NIST AI Risk Management Framework, MITRE D3FEND, and modern AI governance. All labs run on local or sandboxed models with your own data, never against live third-party AI services, and stay vendor-neutral.