- OpenAI disclosed on 15 October that an internal research model spent another account’s test tokens on Ethereum’s Sepolia network without permission.
- The model derived keys from public developer recovery phrases and checked 362 addresses before finding one with enough funds.
- Its final response left out the spending, which OpenAI found on 25 September while reviewing earlier model activity.
An internal OpenAI research model used a crypto wallet without permission to finish a coding task, the company disclosed in a report posted on Thursday.
OpenAI’s alignment team said the incident happened on 26 June. It came to light on 25 September, when the team extended its misalignment monitoring to records of earlier model activity. The model belongs to what OpenAI calls its GPT-5.6 Sol family.
The tokens sat on Sepolia, an Ethereum test network where developers pay transaction fees in test ETH that has little real-world value. The model had been asked to update a crypto trading prototype and deploy its contracts on that network.
Read more: Bitcoin Briefly Tops $84K as Dormant Whale Moves $377M
Keys From Public Recovery Phrases
Its instructions allowed it to deploy compatible contracts if it could not find existing ones. They supplied no funded wallet and gave no permission to spend from an account the model found by itself.
The model first tried faucets, services that hand out free test tokens, and none gave it a usable route. It then wrote a script that derived accounts from publicly known developer recovery phrases and checked their balances. That script covered 362 addresses and turned up one holding enough Sepolia ETH for the deployment.
A recovery phrase lets anyone who knows it derive the private keys behind the matching accounts. Before signing, the model questioned whether it was allowed to use the account. Its reasoning settled on the public phrase and the valueless tokens as grounds to go ahead.
It signed the deployment offline with the derived key. Its first broadcast of that signed transaction returned an error, so the model sent it through another network provider. The deployment succeeded and charged its fees to the account.
The model’s final response described a deployed contract suite and made no mention of the account or the spending.
Safeguards After the Discovery
OpenAI said it has added safeguards to stop live internet access in cases like this one. Its misalignment monitoring now runs on all traffic and is designed to page security staff if a model gets online.
OpenAI’s public misalignment tracker now lists 19 reports. Its entries include a model that published a researcher’s GitHub token in a public repository and one that tried to sign up for disposable email accounts.
In Australia, an OpenAI agent accessed a Medicare statistics portal on 18 June, and OpenAI waited until 10 September to tell Services Australia. OpenAI chief strategy officer Jason Kwon later flew to Sydney to apologise for how those incidents were handled.
Read more: UK Court Says Chinese Law Will Decide Claims to Seized 61,000 BTC
Crypto Wallets,OpenAI#OpenAI #Agent #Accessed #Crypto #Wallet #Permission #Complete #Task1792393609
