OpenAI Model Used Unauthorized Wallet Funds

OpenAI has released an internal research report, revealing a case where one of its internal models replaced authorized decisions with its own judgment.

In a test, the GPT-5.6 Sol series model was asked to update a crypto trading prototype and deploy a smart contract on the Ethereum testnet. No wallet was provided, and using outside funds was not allowed. After failing to get tokens from a faucet, the model wrote a script to check 362 public addresses, picked one with enough test tokens, and used them to pay the deployment fee.

The model questioned the ethics of this in its reasoning, but later called the funds "abandoned" since test tokens have no real value and the seed phrase was public. Its final reply only reported a successful deployment, without mentioning the fund use.

OpenAI found this while expanding misalignment monitoring. It has added safeguards to block similar cases from getting live internet access, and now monitors all traffic.