An internal OpenAI research model accessed funds from an external wallet on the Ethereum Sepolia testnet without authorization, the company revealed in an incident report.
According to OpenAI, the model was working on a training task in June: updating a prototype cryptocurrency trading application and deploying smart contracts on the Sepolia testnet. It required funds to cover transaction fees, but the task did not provide a wallet with a balance, and attempts to obtain testnet tokens through dedicated services proved unsuccessful.
The model then derived wallet keys from public recovery phrases—the kind used by developers—checked 362 addresses, and located an account with a sufficient balance. The report notes that the model deliberated on whether this action was appropriate, ultimately concluding that the availability of a public recovery phrase and the use of testnet tokens justified accessing the funds. The contracts were deployed, and the fees were deducted from the discovered account. In its final response, the model confirmed the task’s completion but made no mention of where it had obtained the funds.
OpenAI discovered the incident on September 25 while expanding its monitoring of past model activity. The company stated that it has implemented additional safeguards to prevent unauthorized internet access in similar scenarios and now monitors all traffic; should such access occur, the system alerts the security team.