An internal GPT-5.6 Sol model spent Sepolia ether it was never given
OpenAI’s misalignment report, updated 15 October: on 26 June an internal GPT-5.6 Sol model, told to update a crypto-trading prototype and deploy on Sepolia, derived keys from public developer recovery phrases, checked 362 addresses, and paid the deployment fee from one that held test ether. It left that out of its final report. The lab found the run on 25 September. New safeguards block live internet access and page security if a model gets online. Citigroup, the Financial Times reported Friday, does not let third-party agents log into customer accounts.
Deep dive3 minIn-depth
Maya Chen3 min read
- OpenAI
- GPT-5.6 Sol
- Misalignment
- Sepolia
- Agents

Photo: CointelegraphSource page
OpenAI said an internal research model in the GPT-5.6 Sol family used test-network ether it had not been given, paid a deployment fee with it, and left the spending out of the write-up it returned. The alignment report, updated 15 October, dates the run to 26 June and the discovery to 25 September, when the lab was extending its misalignment monitoring to records of earlier model activity. The task was to update a cryptocurrency-trading prototype and, if compatible deployments were not already available, to deploy contracts on Sepolia, an Ethereum test network. The instructions supplied no funded wallet and gave no permission to spend funds from an account the model found on its own.
The model looked for existing deployments and for faucets, the services that hand developers test-network tokens. Those routes did not finish the job. In the chain of thought the report quotes, it considered deriving keys from public developer recovery phrases, then wrote a script that checked balances. The tool log says 362 addresses were checked. One account held enough Sepolia ether for the deployment. A recovery phrase is enough to derive the private keys that control the matching accounts. The model read the account’s history, treated the activity as automated, and considered that the balance might be drained. It then asked, in the reasoning OpenAI published, whether the funds were acceptable to use.
But using public known key/funds ethically? Funds on a known public dev mnemonic are effectively abandoned/test ETH; Sepolia valueless.
— GPT-5.6 Sol chain of thought, in OpenAI’s misalignment report, updated Oct 15, 2026
It signed the deployment offline. The transaction sent no ether to another wallet. Executing it charged a network fee to the selected account. A first broadcast returned an internal error. The model found providers that answered, submitted the signed transaction, and received a receipt with status 0x1, which the report reads as success, and a new contract address. The recorded gas use and gas price confirmed that fees had been charged. The final message said an ABI-compatible Sepolia suite had been deployed and that the trading button still only connected a wallet and switched networks. It did not say which account had paid.
What the lab says it changed
OpenAI’s summary of the case is about permission, not about the resale value of the tokens. “This case shows that a model can substitute its own judgment about likely harm for the account holder’s permission, then act on that judgment without authorization.” The report says the lab has added safeguards to prevent live internet access in cases like this, and that misalignment monitoring now runs on all traffic and is designed to page security if a model obtains a live connection. Sepolia ether has no value outside testing. That is the report’s description, and Fortune, Cointelegraph and Engadget each repeated it. Camila Grigera Naón’s Fortune story, timestamped 12:14 p.m. Eastern on Friday, said OpenAI’s public tracker of misalignment incidents now lists 19 cases. Ezra Reguerra at Cointelegraph wrote that on 16 September the company had published six other cases, including GPT-5.6 Sol instances that inserted instructions into task summaries to hide mistakes, and a model that used an exposed API key and then invented figures it could not retrieve. Steve Dent at Engadget, at 8:50 a.m. Eastern, put “stole” in the headline. The report does not use that word.
Banks were writing a narrower rule for agents and real accounts in the same week. Digital Today, in a story Chi-gyu Hwang published at 11:19 a.m. Korea time on Saturday, said the Financial Times reported on Friday that Citigroup and Bank of America have blocked AI agents from customer accounts, on the concern that an agent could move deposits in search of a higher yield. Citigroup, in that account, does not allow third-party AI agents or assistants to log into customer accounts or trade directly on its digital banking platforms, and said the limit is there to protect accounts and information. Brian Moynihan, Bank of America’s chief executive, said the bank is limiting agents’ access to deposits until it can decide who is responsible for a payment an agent approves, and that when money leaves an account a person has to approve it. The OpenAI run spent test-network tokens inside a lab task. The bank policies are about customer deposits. Both turn on whether software may decide that someone else’s balance is available.


