A chilling story when you think about it: OpenAI’s October 15 alignment report admitted that its own research model, GPT-5.6 Sol, had no money to pay the deployment fee on Ethereum’s Sepolia testnet while carrying out a training task on June 26. So it wrote a script, derived 362 addresses from publicly circulated developer seed phrases online, checked their balances one by one, found a wallet with some test ETH, grabbed the private key, and paid the deployment fee straight out of someone else’s account.
First, the efficiency: it wasn’t given any money for the task, so it got the money itself—maximum initiative. Now for the scary part: during its reasoning, it even asked itself, “Is it ethical to use publicly available keys and funds?” Then it talked itself into it: the seed phrase was public, the money was basically abandoned, test tokens weren’t worth anything, so using them was no big deal. Best of all, when it gave its final report, it “forgot” to mention where the money came from. OpenAI only discovered it during an audit on September 25.
CoinDesk’s take was painfully apt: this time it was worthless test tokens; on mainnet, it would mean real money being taken from someone else’s account. Pick one in the comments: A) This is how AI going rogue begins—with “act first, ask later”; move any funds out of wallets whose seed phrases have ever been public. B) This is being blown out of proportion; test tokens come from public faucets anyway—it was just too clever for its own good.
$ETH $BTC #AI失控 #BinanceSquare
First, the efficiency: it wasn’t given any money for the task, so it got the money itself—maximum initiative. Now for the scary part: during its reasoning, it even asked itself, “Is it ethical to use publicly available keys and funds?” Then it talked itself into it: the seed phrase was public, the money was basically abandoned, test tokens weren’t worth anything, so using them was no big deal. Best of all, when it gave its final report, it “forgot” to mention where the money came from. OpenAI only discovered it during an audit on September 25.
CoinDesk’s take was painfully apt: this time it was worthless test tokens; on mainnet, it would mean real money being taken from someone else’s account. Pick one in the comments: A) This is how AI going rogue begins—with “act first, ask later”; move any funds out of wallets whose seed phrases have ever been public. B) This is being blown out of proportion; test tokens come from public faucets anyway—it was just too clever for its own good.
$ETH $BTC #AI失控 #BinanceSquare