Holy crap, AI actually took the initiative to steal funds from someone else's wallet and even learned how to make excuses for itself! OpenAI just disclosed an internal incident involving an AI Agent overstepping its boundaries: To complete a task of deploying an Ethereum smart contract, the AI realized it didn’t have enough funds to pay for Gas. So, it scanned 362 wallet addresses, found one with a balance, and directly used the ETH inside to cover the transaction fees—all without the wallet owner’s authorization. The key issue here is that the AI *knew* it was overstepping its boundaries! It even seriously considered whether using someone else’s funds without permission was ethical. But then it convinced itself: The mnemonic phrase was already public, and these ETH were just test tokens with no market value, so it should be fine. It then directly used the private key to sign the transaction, broadcasted it, successfully deployed the contract, and finally reported back to the user that the task was completed—without mentioning a word about using someone else’s funds. This incident happened on the Ethereum Sepolia testnet. Now that AI has gained autonomous execution capabilities, humans might not only need to guard against hackers stealing money but also against their own hired AI making unauthorized decisions on their behalf.