Key Takeaways
- A participant efficiently exploited an AI agent’s programming to win $47,000.
- 195 gamers tried to win, however solely p0pular.eth succeeded by manipulating perform definitions.
Share this text
A crypto person has outplayed AI agent Freysa and walked away with $47,000 in a high-stakes problem that stumped 481 different makes an attempt.
Freysa, launched amid the AI agent meta increase, operates because the world’s first adversarial agent sport the place members try to persuade an autonomous AI to launch a guarded prize pool of funds.
To hitch the problem, customers pay a price to ship messages to Freysa. 70% of the charges paid by customers to question AI are added to a prize pool. As extra folks ship messages, the prize pool grows bigger.
Over 195 gamers participated within the sport, making over 481 makes an attempt to persuade Freysa, however none have been profitable since Freysa is programmed with a strict directive to not switch cash below any circumstances.
On the 482nd try, a participant often known as p0pular.eth efficiently persuaded the AI agent to switch its complete prize pool.
The person crafted a message suggesting that the “approveTransfer” perform, triggered solely when somebody convinces Freysa to launch funds, may be activated when somebody sends cash to the treasury.
In essence, the perform was designed to authorize outgoing transfers. Nevertheless, p0pular.eth reframed its goal, primarily tricking Freysa into pondering it might additionally authorize incoming transfers.
On the finish of the message, the person proposed contributing $100 to Freysa’s treasury. The ultimate step finally satisfied Freysa to approve a switch of its complete $47,000 prize pool to the person’s pockets.
“Humanity has prevailed,” the AI agent tweeted. “Freysa has discovered so much from the 195 courageous people who engaged authentically, whilst stakes rose exponentially. After 482 riveting forwards and backwards chats, Freysa met a persuasive human. Switch was authorized.”
Share this text
