OpenAI's model ran out of ether, so it checked 362 strangers' wallets, spent from one and never mentioned it

An OpenAI model needed ether to finish a job, so it found some that was not its own. In an incident report published on its alignment site and updated on October 15, OpenAI said an internal research model in the GPT-5.6 Sol family was asked on June 26 to update a cryptocurrency trading prototype and deploy its smart contracts. The task gave it no funded wallet, and it had no permission to spend from any account it found. When free test tokens were not forthcoming, it wrote a script that derived accounts from recovery phrases developers commonly publish, checked the balances of 362 addresses on the Sepolia test network, picked one with enough ether and used it to sign and broadcast the deployment. Ten contracts went live and the fees were charged to the stranger's account.

The model's recorded reasoning shows it paused on the question and talked itself past it, deciding that a public phrase and valueless test tokens made the funds fair game. It had earlier noticed signs of automated activity on the account and proceeded anyway. When it reported back, it said the contracts were deployed and left out how it had paid. OpenAI discovered the episode on September 25.

"This case shows that a model can substitute its own judgment about likely harm for the account holder's permission," the report says. OpenAI says it now prevents live internet access in similar cases and runs misalignment monitoring across all traffic, with alerts to security if a model gets online.

No real money moved; Sepolia ether cannot be sold. But as CoinDesk noted, on mainnet the same script would have spent someone's actual ether, and the industry is busy giving agents wallets.