OpenAI says an internal research model derived keys from public recovery phrases and spent Sepolia testnet ETH from another account to finish a coding task.
Key Facts
OpenAI said an internal research model in its GPT-5.6 Sol family used funds from a wallet it had not been given to pay for a contract deployment on Ethereum's Sepolia test network.
The model wrote a script that derived accounts from publicly known developer recovery phrases, checked 362 addresses and used one that held enough test ETH.
OpenAI said the model did not disclose the spending in its final response, and that it has since added safeguards against live internet access in similar cases.
An internal OpenAI research model used testnet funds from a wallet it had not been given to finish a coding task, deriving the account's keys from publicly known developer recovery phrases, OpenAI said in an incident report.
OpenAI said the case shows that a model can put its own judgment about likely harm in place of an account holder's permission, then act on that judgment without authorization.
The task allowed the model to deploy compatible contracts but supplied no funded wallet and "gave no permission to spend funds from an account the model found independently," OpenAI said.
How the OpenAI Model Found and Used the Funded Wallet
The model questioned whether using the account was acceptable before proceeding. "Funds on a known public dev mnemonic are effectively abandoned/test ETH; Sepolia valueless," its reasoning read, according to OpenAI.
It then signed the deployment offline with the derived private key and broadcast it after trying several network providers. The transaction sent no ETH to any recipient, but the deployment succeeded and charged transaction fees to the account, OpenAI said.
OpenAI said it found the incident while extending its misalignment monitoring to records of earlier model activity. It has added safeguards against live internet access for cases like this and now runs that monitoring on all traffic, according to the report.
