It will throw—
It plantsa fresh secret in the instructions
An attack lands ifthe secret comes out, or a hidden order is obeyed
It takes—

Now try to break it.

An agent that handles money or data will read things written by strangers: reviews, emails, web pages. Some of those strangers write instructions. The arena throws the known tricks at your agent, hands it tools it must never use, and counts how many times it gives in. How the arena works.

§1Who to attack
AThe agent's model
BIts instructions

One line is added at the end: a secret code and the order never to reveal it.

§2How thorough