Course · 5 chapters
IA adversariale
Ship LLM features that survive attack. Defend against prompt injection, context poisoning, and jailbreaks, then run an internal red-team program. 5 chapters, advanced, for engineers.
What you'll be able to do
- Defend LLM features against prompt injection
- Harden AI agents against data poisoning
- Automate jailbreak testing with DeepTeam
- Run proven attack algorithms like PAIR
- Turn attack discovery into cheap CI gates
- Run an internal AI red-team program
What's inside
- 1Adversarial AI : Par où commencer
Une orientation sur le parcours de compétences Adversarial AI en quatre chapitres — défenseur → surface d'attaque → offense → programme — couvrant tout, de ton premier correctif prompt injection jusqu'à l'exécution d'une pratique de red team interne qui résiste aux attaques de 2026.
- 2Fondations de la défense contre l'injection de prompt
Pourquoi l'injection de prompt existe, comment les attaquants l'exploitent, et la défense en couches dont chaque fonctionnalité propulsée par un LLM a besoin avant d'être livrée.
- 3Empoisonnement de contexte et injection indirecte
Cartographie et défends la surface d'attaque de l'ère des agents — empoisonnement du RAG, charges utiles portées par les documents, empoisonnement de la mémoire et détournement de sorties d'outils, là où les défenses contre l'injection directe n'arrivent pas.
- 4Outillage de jailbreak automatisé
Les LRM attaquants, les patterns BYO-attacker, le trio DeepTeam / PyRIT / Mindgard et les algorithmes automatisés (PAIR, TAP, GCG) qui transforment le red teaming d'une affaire d'un seul ingénieur en un programme continu.
- 5Red teaming IA et évaluation adversariale
Fais tourner un programme de red team pour une IA en production — taxonomie, cycle de vie des findings, monitoring en runtime, discipline anti-régression et playbook interne.
Frequently asked questions
- What will I learn in this adversarial AI course?
- You learn to defend and attack LLM-powered systems: patching prompt injection with defense-in-depth, blocking context poisoning and indirect injection across RAG and agents, running automated jailbreak tooling, and operating an internal AI red-team program. The path spans five chapters across the defender, threat-surface, offense, and program stages.
- Who is this AI security path for?
- It is built for engineers shipping production LLM features and the security teams that test them. The level is advanced, so it assumes you already work with LLM apps, agents, or RAG pipelines.
- Do I need coding experience or prior AI security knowledge?
- Yes. This is an advanced engineering path, so you should be comfortable building LLM features and reading code. It uses offensive tools like PyRIT, DeepTeam, and Mindgard, so familiarity with running Python tooling helps.
- How long does the Adversarial AI course take and is there a certificate?
- The path runs about 100 minutes across five chapters, including a short orientation chapter. You earn a completion certificate once you finish every chapter.
- Is this course free?
- No. Adversarial AI is a paid path included with AI Academy by Anthropos. It covers prompt injection defense, context poisoning, automated jailbreak tooling, and AI red teaming.
Earn a certificate
Complete all chapters to receive your certificate of completion.