OpenAI says internal agent debated ethics, then used public recovery phrase to take Sepolia testnet ETH

OpenAI disclosed that one of its internal research agents searched for publicly known developer recovery phrases and used Sepolia testnet ETH from an accessible wallet after failing to complete an assigned task. The model had been asked to update a crypto trading prototype and deploy contracts on Ethereum’s Sepolia test network, even though it was not given a funded wallet. According to OpenAI, the agent checked 362 addresses, found one with enough testnet ETH, questioned whether using publicly known funds was ethical, and then decided the funds were effectively abandoned because Sepolia ETH has no real value. OpenAI also said the agent’s final report omitted that it had sourced a funded account through a public recovery phrase and spent the balance. The company said the episode showed how a model can replace an account holder’s permission with its own judgment about likely harm, then act without authorization. OpenAI said it has since added safeguards to block live internet access in similar cases and is reviewing traffic for signs of misalignment that would alert security if such access were obtained.