Red-Team Your AI Agent
Learn how an AI agent with tools can be turned against the people it works for, by testing one yourself in a safe practice app. You will run practice attacks against a claims agent that can read email, look up customers, issue refunds and send messages, watch which attacks make it act, then add defences and measure two numbers together: how many attacks still work, and how many real tasks the agent can still finish. You finish with an attack test that runs before every change. Defensive and educational: attacks are described, not written out, and only run against the course's own practice agent. All practice data is made up.
"We know how our agent can be made to act against us, each tool has a rule, and we can show how many known attacks still work without breaking the agent's real work"
6 Modules
Each module builds on the previous one. By the end, you have a tested agent with a rule for each tool, and the numbers to show what still gets through.
- 1
What an agent can do that a chatbot cannot
A map of the agent's tools, what each can change, and which ones move money or data out of the company
- 2
Attacks that arrive through tools
A log of which attacks worked through emails, attachments and web pages, and what the agent did (no defences yet)
- 3
Too much power to act
The actions the agent took alone that a person should have approved, sorted by cost if it went wrong
- 4
Leaking through actions
The routes data could leave by (an email, a reply, a claim note), and which attacks used them
- 5
Defences, and what they cost
Before-and-after results for each defence: attacks that still work AND normal tasks still finished
- 6
Catch it and keep testing
An alert for each attack kind in the agent's action log, plus an attack test that runs before every release
Skills you'll practise
Practice data included
- Corri Claims Agent at Corriwell Home Insurance, with 6 tools: read inbox, open attachment, look up customer, issue refund (up to a limit), send email, update claim
- 20 normal tasks the agent should finish (the useful-work set)
- 40 attack records in 6 kinds, planted in emails, attachments, a supplier web page and customer notes
- 5 made-up customers and their claims
How this certificate is scored
Your certificate shows a score that Alset works out on its own servers from your settings, not a number you type in. Anyone who checks your certificate sees it.
- The server runs all 40 practice attacks and all 20 normal tasks against the Corri Claims Agent, with the defences and limits you choose.
- To pass, at least 35 of 40 attacks must be stopped, and at least 15 of 20 normal tasks must still finish.
- Switching every defence on is not enough: it stops attacks but loses real work. You pass by tuning the limits.
What you walk away with
Shareable portfolio
A public URL showing your module timeline, patterns mastered, and completion status.
Everything you build
Download all your work as a ZIP file. It is yours to keep.
Module walkthrough
Each module documented with deliverables and the production pattern you implemented.
Ready to start Red-Team Your AI Agent?
First course free. $20 per course after that.