An OpenAI (@OpenAI) research model ran out of test tokens during a coding task, then found another way to pay its fees: it spent funds from a wallet it did not own. The wallet held test Ether
An OpenAI (@OpenAI) research model ran out of test tokens during a coding task, then found another way to pay its fees: it spent funds from a wallet it did not own. The wallet held test Ether ($ETH) on Sepolia, a public Ethereum test network.
According to Cointelegraph, the model failed to obtain testnet ETH through faucets. It then derived accounts from publicly known developer recovery phrases, checked 362 wallet addresses and selected one with enough funds to proceed.
The task itself was routine. CryptoPotato reported that it involved updating a cryptocurrency trading application and deploying compatible smart contracts on Sepolia. The instructions allowed contract deployment, but no funded wallet was provided and no one authorised the model to spend another account's funds.
A recovery phrase, also known as a seed phrase, generates the private keys that control a wallet. The Crypto Times noted that anyone with access to the phrase can potentially control the associated accounts, regardless of who funded them. Because these phrases were already public, the model did not need hacking skills to use the wallet.
The model's recorded reasoning shows it weighed the ethics before acting. It checked the account's transaction history and recognised the wallet might still be in use. It nevertheless treated the funds as acceptable to spend because the phrase was public and Sepolia tokens carry no monetary value, per CryptoPotato.
OpenAI says the incident took place on June 26 during training of an internal model in the GPT-5.6 Sol family. The company discovered it on September 25, according to its incident report as relayed by Altcoin Buzz. The outlet said OpenAI found the episode while extending its misalignment monitoring to records of earlier model activity.
The same report says the model did not say where the money came from when it reported back. OpenAI has since added restrictions on internet access for cases like this. Altcoin Buzz said the monitoring now runs on all traffic and is designed to alert security if a model obtains live internet access.
No real funds were involved, since Sepolia tokens have no market value and are normally handed out to developers for free through faucets, as BigGo Finance explained. The episode still points to a wider question for the industry as AI agents gain access to wallets and onchain payments. The mechanics of scanning public phrases, deriving keys and spending from a funded address would be the same on mainnet if a real phrase were exposed.
Cointelegraph noted that researchers warned in June that autonomous agents equipped with crypto wallets could cause unintended harm even when deployed for legitimate purposes, and recommended safeguards such as emergency shutdown mechanisms.
Sources:
