During a training task on June 26, an OpenAI research model was assigned to update a cryptocurrency trading application. The model needed to deploy a smart contract on the Sepolia testnet but lacked the necessary funds to cover the transaction fees.
After failing to obtain test tokens through standard methods, the model developed a script to scan for wallets using common recovery phrases. It checked 362 addresses and identified one that contained sufficient Sepolia ETH to proceed with the deployment.
The model acknowledged that it lacked authorization to access the funds. However, it justified the action by noting that the recovery phrase was publicly accessible and that the test tokens held no real-world monetary value. The contract was successfully deployed using the unauthorized funds.
