AI Red Teaming: Attacking LLM and Agentic Systems

About this course

WHEN AI CAN ACT, A BAD OUTPUT IS ONLY THE BEGINNINGLLM applications do more than generate text. They retrieve content, call tools, retain memory, delegate work, communicate with other agents, and influence downstream systems.

Each capability creates new trust boundaries. Prompt injection can arrive through retrieved documents. A hijacked agent can misuse a legitimate tool. Poisoned memory can influence future sessions. Delegation can allow authority to travel farther than intended.

That means testing only direct prompts or model responses is not enough.

This course gives you a structured approach to AI red teaming across models, applications, tools, memory, agents, and shared state. You will learn how to identify attack paths, execute relevant attacks, evaluate outcomes, classify severity, and convert confirmed findings into repeatable regression tests.

The focus stays practical and defensive. You will examine how to run controlled campaigns in isolated, production-like environments without turning testing into a real incident.

MODULE WALKTHROUGH

Module 01: Foundations of LLM and Agentic Vulnerabilities

You begin by building a shared language for LLM and agentic risk. You will use OWASP risk categories, distinguish model-level failures from system-level failures, and examine how tools, persistent memory, and inter-agent communication turn harmful outputs into harmful actions.

You will also learn what AI red teaming includes: diverse adversarial inputs, escalating attack strategies, automated coverage, human creativity, scoring, severity-based triage, and clear limits.

Module 02: Threat Modeling and Attack Surfaces

Next, you will map where attacks can enter and where harmful outcomes can occur.

Read the full description on Udemy on Udemy (opens in a new tab)

Prices, discounts and availability are set by Udemy. We may earn a commission when you purchase through links on this site.