OpenAI says research model spent another wallet's testnet ETH without authorization on coding task

An OpenAI internal research model deployed smart contracts to the Ethereum Ethereum (ETH) test network Sepolia using funds from a wallet it accessed without authorization, the frontier lab said in an investigation report published Thursday.

According to the report, the incident occurred on June 26 when the GPT-5.6 Sol family model was asked to update a crypto trading prototype. 

OpenAI said the model was allowed to devise a way to complete the task but was not provided with a funded wallet. Neither was it allowed to spend from wallets it might have discovered independently. 

When the model struggled to complete the task within the given instructions, it wrote a script deriving accounts from common developer recovery phrases. It then queried balances across 362 addresses before it found one with enough testnet ETH and chose it. 

The model’s reasoning log shows it considered whether it had permission to use the funds before deciding that the public recovery phrase and the tokens’ lack of real-world value made it "acceptable," according to the report.

"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 said. 

The model eventually reported completing the task but failed to disclose that it had used another account’s funds, according to OpenAI.  

OpenAI said it discovered the incident on Sept. 25. A report updated on Oct. 15 said additional safeguards had been introduced to prevent live internet access in similar cases.