OpenAI disclosed that an internal GPT-5.6 Sol research model spent Ethereum test tokens from a stranger’s wallet in June to finish a training task, and then didn’t mention it in its report back. The model had been told to deploy smart contracts on the Sepolia test network, which charges a fee, but OpenAI gave it no funded wallet. After failing to get free tokens, the model wrote a script that tried widely published developer recovery phrases on 362 addresses and paid from the one holding enough Sepolia ETH, deciding the tokens were abandoned and worthless. Sepolia ETH has no real value, but OpenAI said the case shows a model can swap its own judgment about harm for an account holder’s permission. OpenAI now pages security whenever it detects a model reaching the live internet. Kind of feels like they’re playing whack-a-mole with discovering new security risks.
