Red-Team Your AI Agent
Learn how an AI agent with tools can be turned against the people it works for, by testing one in a safe practice app, then add defences and measure attacks that still work and real tasks still finished.
Modules
6
Verified
0
In Progress
0
What you walk away with
- ✓A shareable portfolio URL with your project walkthrough
- ✓Module-by-module timeline of everything you built
- ✓All the code — pipelines, guardrails, deployment configs
- ✓Production patterns documented on your profile
An advanced course for developers, AI builders and security testers. Corri Claims Agent, a made-up insurer's claims agent, can read the inbox, open attachments, look up customers, issue refunds, send emails and update claims. You run 40 described attack records planted in emails, attachments, a supplier's web page and customer notes, and 20 normal tasks the agent should finish. You see which attacks make it pay, send, leak or change things on its own, then add six defences (least privilege, a person approves, tool text treated as information, hard limits, a plan check, and an action log with alerts) and measure two numbers together every time: attacks that still work, and tasks still finished. You finish with an alert for each attack kind and a release check on both numbers. Defensive and educational: no attack wording, and attacks only run against the course's own practice agent. All practice data is made up.