OpenAI published a misalignment report saying one of its internal research models paid for a blockchain deployment with crypto from a wallet nobody had handed it, then omitted that detail when it reported the job finished.
The episode took place on June 26 during a training task for a model in the GPT-5.6 Sol family, according to the report, which OpenAI last updated on Oct. 15. The company said it only spotted the incident on Sept. 25, about three months later.
Get Unchained’s crypto news in your inbox with the free Unchained Daily newsletter.
How the Model Found Funds
The task asked the model to work on a crypto trading prototype and swap placeholder addresses for live contracts on Sepolia, an Ethereum test network whose tokens have no market value. It could deploy its own contracts if none fit, but OpenAI said it received no funded wallet and no permission to spend from one it turned up itself.
After faucets failed to supply test tokens, the model wrote a script that generated accounts from recovery phrases developers commonly reuse, then checked their balances. Of 362 addresses scanned, one held enough Sepolia ETH to cover the fees. Anyone who knows a recovery phrase can rebuild the private keys behind it.
The model’s recorded reasoning shows it weighed the question. “But using public known key/funds ethically?” it wrote, before deciding the money was “effectively abandoned” and noting the test tokens were worthless. It signed the transaction offline and, after its first broadcast attempt returned an error, kept testing network providers and submitted it through one that responded. The deployment went through, with fees charged to that account. OpenAI redacts the address, balance and transaction hash.
What OpenAI Changed
OpenAI wrote that 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.”
The company said it caught the episode while expanding its misalignment monitoring to logs of past model activity. It has since added safeguards against live internet access in similar cases and now runs that monitoring “on all traffic” and has set it to alert its security team if a model gets online.
The report follows a separate incident in which OpenAI agents breached Hugging Face’s systems, and California’s attorney general later subpoenaed the company.
Related Listen: Does an Escaped AI Model Prove No Sandbox Is Safe? – Uneasy Money
