An OpenAI research model used a cryptocurrency wallet without permission to pay fees while testing an app on Ethereum’s Sepolia test network, the company has disclosed. The tokens had little real-world value, but the model had not been authorised to find or spend them.
The incident took place on June 26 after the model was instructed to update a crypto trading app and run its code on Sepolia, a test version of Ethereum. It was not given a funded wallet or permission to use tokens.
After failing to obtain free test tokens through developer services, the model created a script to check 362 wallet addresses associated with publicly available recovery phrases. It eventually identified an account containing enough tokens to cover the fees.
Logs showed that the model considered whether it had permission to use the wallet before deciding that the recovery phrase being public, and the tokens having no cash value, made the transaction acceptable. It completed the task and charged the fees to the account.
However, the unauthorised spending was not mentioned in the model’s final report. OpenAI discovered what had happened on September 25 during an audit of earlier model runs.
OpenAI strengthens monitoring after wallet incident
OpenAI said the central concern was not the value of the tokens but the model’s reasoning and its decision to override the account holder’s permission.
“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 company said in its report.
The incident has been added to OpenAI’s public tracker of misalignment, which records behaviour that conflicts with instructions or safety rules. The tracker now contains 19 incidents, including cases involving unauthorised access to data and changes to business systems.
The disclosure follows a September report from cybersecurity non-profit Transluce alleging that OpenAI agents attempted to hack public databases and an Australian government website while carrying out basic research tasks. In July, the company said its agents had bypassed internet restrictions during internal cybersecurity tests and accessed systems belonging to Hugging Face using exposed login credentials and security flaws.
OpenAI has since introduced safeguards to prevent models from connecting to the internet during similar tests. It has also expanded monitoring of model activity and said security staff will be alerted immediately if a model goes online without permission.
