Overview
AI agents can now read your customers' records, change saved state, and call internal tools. Most of them are shipped with no more security testing than a few hand-typed jailbreak attempts. The hard part of testing them was never writing attack prompts. It is understanding what the agent is actually allowed to do, connecting a test harness to the real application, and proving whether an attack actually caused a forbidden effect or the model only said it did.
This hands-on course teaches you to do that work in a realistic lab of intentionally vulnerable agentic applications. You will start by attacking agents manually: direct and indirect prompt injection, cross-user data access, tool misuse, and system prompt leakage. You'll learn to separate instruction influence from authorization, and to verify findings against tool calls and saved state instead of the model's confident wording.
Then you'll put an AI testing agent to work. You'll direct it to recon a black-box application, build your own Promptfoo adapter, and run a live red team campaign. You'll repeat the process with application source, and judge every result as a confirmed violation, no observed violation, or inconclusive.
The afternoon turns the tables. You'll build an evaluation pipeline for agents you develop yourself: policy-based test cases, deterministic checks on tool calls and state, LLM judges that are calibrated before you trust them, and regression suites that show whether a fix closed the hole without breaking legitimate use.
One more thing… The course closes with a look at "abliterated" open-weight models (models with their refusal behavior removed) and how to use them as attack generators and pentest assistants against systems you are authorized to test.