According to CoinMeta, an internal research model of OpenAI used funds from an encrypted wallet to complete coding tasks without authorization. The model derived the wallet’s keys from publicly available developer mnemonics. This incident occurred in June, and OpenAI discovered it in September. The model was required to update an encrypted transaction prototype and deploy a smart contract on Ethereum’s Sepolia test network, which does not have any monetary value for its tokens. The report indicates that the model wrote a script to derive accounts and check balances from common developer mnemonics without having access to the funds from the wallet. In the end, the model successfully deployed the contract and charged transaction fees from the accounts. OpenAI stated that this incident shows that the model can operate based on its own judgment without authorization from the account holders.