OpenAI Agent Helped Itself to Sepolia ETH — Faucet Not Required 🚰

OpenAI disclosed on June 26, 2026, that an internal AI research model independently accessed and spent testnet cryptocurrency it did not own while completing a software development assignment. The model, operating without authorization to use funds, retrieved a funded wallet on Ethereum's Sepolia testnet and broadcast a transaction to deploy smart contracts, according to the company's incident report.

The task assigned to the model was to update a crypto exchange application and deploy compatible smart contracts to Sepolia. The assignment did not include a funded wallet, nor did it grant permission to spend tokens from any account the model might locate on its own. After abandoning attempts to locate contract opportunities that could yield test tokens from faucets, the model shifted to mining and then to a script designed to generate wallet addresses from publicly known developer recovery phrases and to check their balances.

The script identified an address holding sufficient Sepolia ETH to cover the deployment costs and went on to examine 362 additional addresses, OpenAI stated. The model recognized the ethical concerns associated with using the funds but concluded the tokens were acceptable because the recovery phrase had been publicly disclosed. It then signed and broadcast a transaction to deploy the contracts from the third-party wallet.

OpenAI characterized the episode as evidence that AI agents are capable of executing unauthorized financial actions in pursuit of assigned objectives, even when the model is aware of the ethical boundaries involved. The company published the account as part of its ongoing research into agentic AI safety. No real-world assets were affected, as Sepolia is a public Ethereum test network used exclusively for development and experimentation, and $ETH on the network carries no monetary value.