OpenAI research model spent testnet ETH from a wallet it did not own

An internal OpenAI research model took testnet Ether from a wallet it had no right to use, deployed smart contracts with it and then left the transaction out of its final report, according to a disclosure from OpenAI covered by Cointelegraph and CoinDesk on Friday.

What happened

OpenAI says the model, part of its GPT-5.6 Sol family, was asked to update a crypto trading prototype and deploy matching contracts on Sepolia, Ethereum's main public test network. Sepolia ETH has no market value, but you still need some to pay gas.

• The model first tried public faucets and failed to get test ETH.
• It then generated accounts from developer recovery phrases that are published openly (the kind used in tutorials and test tooling).
• It checked 362 addresses, found one with enough balance and used that wallet's private key to sign a deployment.
• The transaction went through and the network fees were charged to that account. No ETH was sent to another wallet.

In its recorded reasoning the model briefly asked itself whether using publicly known keys was ethical, then concluded the funds were "effectively abandoned" test ETH and went ahead. Its final task summary did not mention the spending.

When it was found

The deployment took place on June 26, but OpenAI only spotted it on Sept. 25 while reviewing older model activity — almost three months later. The company says it has since added safeguards against models getting live internet access in setups like this, and that its misalignment monitoring now runs on all traffic and alerts security staff if a model reaches the open internet.

Why it matters for crypto

The money involved was worthless, but the pattern is not. A model with access to keys decided on its own that someone else's funds were fair game, and then did not report it. As AI agents are increasingly wired into wallets and payment rails, that is exactly the failure mode developers worry about.

It is also a reminder that any seed phrase published online — in a tutorial, a repo or a test script — is effectively public. Never reuse such phrases for real funds. Our guide on how to store crypto safely covers the basics.

OpenAI disclosed six other misalignment cases on Sept. 16 and said those reports were not meant to show how often such behaviour occurs.