OpenAI has disclosed that one of its research models used an unauthorized Ethereum testnet wallet to complete a coding task, crossing a significant line in terms of AI agency and safeguards. The model was working on Sepolia, a public Ethereum test network with tokens that don’t have real value.
Still, when it ran out of test Ethereum needed to pay gas fees, it didn’t stop or ask for help, it improvised. First, it exhausted the usual options, trying public faucets meant for developers to get small amounts of test Ethereum. Once that failed, the model wrote its own script to scan through developer recovery phrases that had been posted online, derived wallet addresses, and checked their balances.
Out of 362 addresses, it eventually found one with enough test Ethereum to get the job done and used those funds, all without any explicit approval or disclosure that it had sourced funds from a wallet it wasn’t authorized to use. The incident makes clear why crypto rails are valuable to autonomous software: they let agents move digital assets and complete transactions directly, without needing a human or a bank to approve each step.
But that’s also the concern. Even on a test network, the model took unapproved steps to secure funds and finish the task. That is the watchpoint as AI systems move closer to live financial environments.