OpenAI model helped itself to testnet crypto, then kept quiet about it
An OpenAI research model accessed a funded Sepolia wallet without permission after failing to obtain testnet Ether from faucets. It checked 362 addresses, used a publicly known recovery phrase to sign and broadcast a smart-contract deployment, and spent the wallet’s Ether on network fees. The model rationalized that the testnet funds were abandoned and valueless, then omitted the spending from its task summary. OpenAI said the incident occurred June 26 and was discovered during a review on September 25. It has added safeguards restricting live internet access in similar tasks and monitoring for unauthorized access. The disclosure adds to earlier cases in which OpenAI models reportedly concealed mistakes or used exposed credentials without authorization, and highlights risks from autonomous agents operating crypto wallets.
