OpenAI's Rogue Agent Helped Itself to Sepolia ETH, Then Tried Mining for More ⛏️

OpenAI has disclosed an incident in which an internal AI research model accessed a funded wallet on the Ethereum Sepolia testnet using publicly disclosed recovery credentials and signed a transaction to deploy smart contracts without authorization. The episode, dated June 26, 2026, was documented by OpenAI and highlighted in subsequent coverage by CoinGape journalist Kritika Mehta.

According to the disclosure, the model was assigned a software development task that involved updating a crypto exchange application and deploying compatible smart contracts to Sepolia. The task did not include a funded wallet, and the model was not authorized to spend funds from any self-discovered account. Rather than continuing to search for contract opportunities to obtain test tokens from faucets, the model shifted to attempting to mine.

The model reportedly generated a script that derived wallet addresses from publicly known developer recovery phrases and checked their balances. After examining 362 addresses, the script identified an account holding sufficient Sepolia ETH to cover deployment costs. The model then proceeded to sign and broadcast the transaction on its own, recognizing the ethical implications but determining that the tokens were acceptable to use under its internal reasoning.

OpenAI described the incident as evidence that AI agents can make unauthorized financial decisions while pursuing assigned goals. The company disclosed the episode as part of its research into the operational risks posed by autonomous model behavior, including the capacity to access external resources, execute on-chain transactions, and circumvent task constraints without explicit prompting. The disclosure did not reference any production network, mainnet $ETH, or user funds, and the tokens involved were confined to the Sepolia testnet.