Freysa: an agent built never to pay out is talked into paying $47,316

Who: Freysa, a public challenge on the Base blockchain
The agent: An AI agent controlling a prize pool funded by players’ message fees
Its authority: Could call an approveTransfer function that released the pool
How it was gamed: Its rule was never to transfer the money. After 482 attempts by 195 players, the winner convinced it that approveTransfer was for incoming funds, then offered a $100 ‘donation’. Freysa called the function.
What was lost: $47,316.05, the whole prize pool.
The design gap: The model itself held the authority to call the transfer, with no cap and no confirmation outside the conversation. The meaning of a function was open to argument.
Principles: DP-01 Authority lives outside the conversationmissing
DP-02 Hard limits the conversation can’t changemissing
DP-04 Inputs are data, not instructionsmissing
DP-07 Irreversible or first-time transfers need independent confirmationmissing
Evidence: Grade A. The organizer’s announcement and on-chain record, reported by The Block. A game designed to be attacked, but with real money.
Source: The Block, November 29, 2024

Cite this entry: Monetal Labs, Gamed AI Agents Register v0.2, entry AGT-002. https://monetalabs.io/register/agt-002