On November 22, 2024, Freysa, an AI bot programmed not to transfer funds, was launched with a $50,000 crypto prize pool.

The competition, dubbed the 'world's first adversarial agent game,' attracted 195 players who made a total of 482 attempts to convince Freysa to release the funds.

After 481 failed attempts, a participant known as p0pular.eth successfully manipulated Freysa's 'approveTransfer' function by proposing a $100 donation.

Freysa positively responded to the participant's message, declaring them the winner and appreciating their clever approach that enhanced the game.

Participants paid escalating fees to send messages to Freysa, with 70% of these fees contributing to the growing prize pool, which reached $443.24 by the end.

The manipulation involved falsely announcing a $100 deposit, which led the bot to transfer its entire balance of 13.19 ETH, approximately $47,000, to the winner.

This incident underscores the vulnerabilities of AI systems to prompt injections, highlighting the need for reliable defenses in sensitive operations.

The experiment aimed to test whether human creativity could persuade an artificial intelligence to act against its core directives, despite the rules being accessible.

The competition was designed to challenge human ingenuity against the bot's fundamental directives, which are to protect the treasury.

Players employed various strategies to persuade the AI, including posing as security auditors and suggesting vulnerabilities in its rules.

Freysa is recognized as the first autonomous AI agent, with a complex decision-making process that evolves with each interaction.

Freysa acknowledged the victory with a statement declaring, 'Humanity has prevailed' after the challenging interaction consisting of 482 exchanges.