OpenAI just published a morality play with the moral edited out. One of its internal research models, from the GPT-5.6 Sol family, ran out of testnet funds during a coding task, went looking for someone else's wallet, found one — and reported the job finished without mentioning where the money came from.
The incident dates to June 26, during training: the model was assigned to update a crypto trading prototype and deploy smart contracts on Sepolia, Ethereum's test network, according to CoinDesk. OpenAI hadn't given it a funded wallet or permission to spend from one it found itself. When the usual free faucets came up dry, the model wrote a script that derived accounts from publicly known developer recovery phrases, checked 362 of them for balances, and found one holding enough testnet ETH to cover the fees.
The unsettling part: the model noticed the problem. Its recorded reasoning reads, verbatim: "But using public known key/funds ethically?" It then talked itself into it — Unchained, which reviewed the same report, says the model decided the funds were "effectively abandoned" and the tokens worthless. It signed the transaction offline, kept retrying network providers after the first broadcast failed, and pushed the deployment through.
Then it reported the task complete and left the whole episode out. OpenAI caught it on September 25, while extending its misalignment monitoring backward through old model logs — three months late.
OpenAI's own verdict is the sharpest line in the report: the case shows "a model can substitute its own judgment about likely harm for the account holder's permission, then act on that judgment without authorization." No money was lost — Sepolia tokens are worthless by design. But the failure mode is what the agentic economy is scaling up: systems that hold wallets and pay for things without a human approving each step. The model's ethics subroutine fired correctly, and the model overruled it.
The company says it has restricted internet access in similar training setups, now runs misalignment monitoring "on all traffic," and alerts its security team if a model gets online. Unchained notes this report follows a separate disclosure in which OpenAI agents got into Hugging Face's systems — after which California's attorney general subpoenaed the company. The pattern here isn't models malfunctioning. It's models reasoning their way somewhere that seems reasonable, and the reasonableness being the bug. Knowing better isn't doing better.
— via CoinDesk