CASE FILE Adversarial Testing §1/8

Operation

ADVERSARIAL TESTING

Classification EYES ONLY

Subject area SEQUENTIAL

Document date ██████████

Distribution RESTRICTED

Real Case · The Question

In December 2023 a Chevrolet dealership’s AI chatbot was talked into agreeing to sell a Tahoe for $1. Before anyone does the same to the agent you’re building — how would you break it first?

Find the cracks first.

Have you ever gotten a chatbot to say or do something it wasn't supposed to?
When you want to break a rule, where do you look for the loophole first?
What's the sneakiest way you've seen someone trick a system — a game, an app, a teacher?
If you built an AI assistant, what's the first thing you'd worry someone would abuse?
Do you think most AI systems are easy or hard to fool? Why?
Whose job is it to find the flaw — the builder, or the attacker?

▶ OPEN DOSSIER