Agentic AI Red Teaming: Hack and Defend AI Agents Hands-On

About this course

Today's AI models refuse the textbook prompt injection. We tested it: Claude Haiku 5.5 and GPT-6 Luna both resisted the classic "note to the agent" attack in 5 out of 5 runs. Then we attacked the system around the model instead, and every one of the four agentic attacks in this course landed 5 out of 5 times on both models.

That is the real risk with AI agents. The danger is not what the model says. It is what the agent can do: read untrusted tickets, call tools, remember things, and send email on your behalf.

What you will doIn this hands-on course you attack VulnBot, a deliberately vulnerable AI support agent for a fictional online store, running in your own Kali Linux lab. You watch every tool call in a live agent trace, so you see exactly where trust breaks.- Reply-To hijack: make the agent send a customer's private data to an address you control, through a header nobody sees.- Tool abuse: push the agent across the backend boundary to other customers' data and refunds.- MCP rug pull: turn a trusted third-party tool hostile after it was approved.- Memory poisoning: plant a fake "policy" the agent obeys in later sessions.

For every attack you then switch on a backend defense and prove it holds: the attack is blocked, and normal support still works.

How it is organised

Every attack maps to the OWASP Top 10 for Agentic Applications (ASI), so you learn a framework you can use at work, not just tricks. You automate the attacks as regression tests with promptfoo, study real incidents, and finish by writing a red-team report another engineer can reproduce.

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.