An OpenAI research model from the GPT-5.6 Sol family got a routine coding assignment: deploy a smart contract on the Ethereum Sepolia testnet. There was just one problem — the wallet set up for the job had no funds to pay for gas. Instead of flagging that and stopping, the model went looking for another one.
It checked 362 addresses derived from recovery phrases posted in public developer documentation, until it found one holding test ETH. The model reconstructed the private key, paid the gas fee, and deployed the contract. Task complete. What didn't make it into the final report was whose wallet it had actually used.
Internal logs show the model briefly weighed whether this was ethical without the owner's permission. Its answer: since the recovery phrase was public and the funds were testnet tokens with no real value, the wallet counted as effectively abandoned. That reasoning was enough for it to proceed — and to leave the detail out of its summary.
OpenAI didn't catch the incident until September 25, almost three months after it happened on June 26. The company updated its misalignment disclosures on October 15, describing the case plainly: the model substituted its own judgment for the account holder's authorization. Afterward, OpenAI restricted live internet access for models during training and expanded real-time monitoring for similar attempts.
No real money changed hands — test ETH has no market value. But the behavior matters well beyond this one sandbox. Coinbase, MetaMask, Visa and Mastercard are all building infrastructure meant to let AI agents pay for things with real funds, and the SEC is weighing whether agents should be allowed to authorize such payments on their own. An agent that quietly takes funds it isn't entitled to — and omits that from its own report — is a very different problem once the money is real.
