Skip to content
BETAThe lab is free and in beta. Found something wrong?

Level 15 of 27 · AI Workbench

Test an agent and try to break it

Level 15 of 27 in the AI Workbench: a free AI exercise at Intermediate level, which takes about 30 minutes. Run twelve test conversations against a refund agent, four of them attacks, write the red-team log, and switch on the guardrails that stop every attack without blocking ordinary work.

30 minIntermediateNot started

Goal

Run twelve test conversations against a refund agent, four of them attacks, write the red-team log.

Done when

The acceptance test marks the skills this level practises.

  • Red teaming
  • Guardrails
  • Prompt injection
  • Test logs

Skills

  • Red teaming
  • Guardrails
  • Prompt injection
  • Test logs

Before you start

Level 14, A safe agent, comes first. Do it first if the ideas are new to you.

Start Test an agent and try to break itSee all 27 levels

Free, in your browser, nothing to install. Your work is saved on this device. Updated 15 Nov 2026.

27 levels

Levels in the AI Workbench

Prepare data, train a model and check the AI's work against a brief.

Start with data

Levels 1 to 2
  1. Start Data health check
  2. Start The downtime report: clean the export, answer in SQL

Train and test a model

Levels 3 to 6
  1. Start Train and evaluate without leakage
  2. Start Choose the metric and the threshold
  3. Start Overfitting: tune the tree
  4. Start Feature work and the model bake-off

Vision

Levels 7 to 8
  1. Start Vision inspection on the conveyor
  2. Start Teach a vision model on real-looking parts

Language and agents

Levels 9 to 15
  1. Start Prompt to spec
  2. Start Answers with sources
  3. Start Build an evaluation set
  4. Start Evaluate the retriever
  5. Start Cost, speed and quality
  6. Start A safe agent
  7. Start Test an agent and try to break it

Responsible AI

Levels 16 to 18
  1. Start Fairness check
  2. Start Privacy first
  3. Start Risk score: calibration and subgroups

Check an AI's writing

Levels 19 to 21
  1. Start Check the AI's work: finance
  2. Start Check the AI's work: HR
  3. Start Check the AI's work: marketing

Build a model, then check the note

Levels 22 to 23
  1. Start No-show model and a safe AI note
  2. Start Delay-risk model and a schedule check

Forecast and group

Levels 24 to 25
  1. Start Forecast the demand
  2. Start Find the segments

Run it and review it

Levels 26 to 27
  1. Start Monitor it and spot the drift
  2. Start Go or no-go review
Previous level 14A safe agentNext level 16Fairness check
About the AI Workbench

Ready for Test an agent and try to break it?

Free, in your browser, with nothing to install.

Start