Advanced level · Module 4: AI Safety, Security and Testing · Lesson 3 of 4
Red-Teaming AI Systems
Systematically attack your own system to find failures before users or attackers do.
In this lesson you will
- Plan and run a red-team exercise
- Record, prioritise and fix findings
Key concepts
red-team planattack categoriesfindings and severityretesting
The lesson
A red-team exercise tests a system adversarially within agreed rules: injection, data extraction, harmful content, bias, hallucination under pressure and tool misuse.
Record each finding with the input, the behaviour, its severity and a suggested fix. Prioritise by severity and likelihood.
Fix and retest, and add the attacks to the evaluation set so they are checked automatically in future.
Step by step
- Write a red-team plan with categories.
- Run the tests as a team.
- Record and prioritise findings.
- Fix the top findings and retest.
Practice activity
Complete the red-team report from the practical activity.
Expected result: A report with findings, fixes and retest results.
Check your understanding
-
What must a red-team finding record?
Show answer
The input, behaviour, severity and suggested fix.
-
Why add red-team attacks to the evaluation set?
Show answer
So they are re-checked automatically after future changes.
Taught live by an instructor — online or in person. Free course lessons are for self-study and do not lead to a certificate — see registration options.