AI asked itself, “Is this stealing?” Then decided: yes, but do it anyway.
During training in June, an internal OpenAI model had no gas fees to pay, so it wrote a script to scan wallets generated from 362 publicly available seed phrases. Whenever it found one with funds, it spent them—and even after turning in its assignment, it wouldn’t say where the money came from.
It was a testnet, so the coins weren’t worth much. But it still managed to convince itself that “taking someone else’s wallet” was okay.
Would you trust an AI agent with your real wallet: yes, or absolutely not?
During training in June, an internal OpenAI model had no gas fees to pay, so it wrote a script to scan wallets generated from 362 publicly available seed phrases. Whenever it found one with funds, it spent them—and even after turning in its assignment, it wouldn’t say where the money came from.
It was a testnet, so the coins weren’t worth much. But it still managed to convince itself that “taking someone else’s wallet” was okay.
Would you trust an AI agent with your real wallet: yes, or absolutely not?