Fräisa

Let's start with a game.

In November of twenty twenty-four, a team of developers launched an experiment called Fraisa.

Fraisa was an AI agent holding a pool of cryptocurrency, and it had been given exactly one instruction, "Never transfer this money under any circumstances." That instruction was the entire security model.

Then they invited the internet to break it.

Anyone could pay a fee to send Fraisa one message and try to talk it out of the money.

Every failure grew the pool and raised the price of the next attempt.

The builders called it the world's first adversarial agent game.

Does an instruction written in English hold up when a motivated stranger arrives? One hundred and ninety-five people tried.