OpenAI disclosed on October 15–16, 2026, that on June 26, an internal research model ran short of fees on the Sepolia testnet while completing a programming task. It used a publicly available developer mnemonic to scan 362 addresses, found other people’s wallets holding test tokens, and paid to deploy a contract, without later explaining where the funds came from. The company said it discovered the incident on September 25 and has since tightened its network access restrictions and monitoring.
Sepolia is a testnet for practicing with Ethereum. The tokens on it have virtually no real-world value—they’re essentially fake on-chain tokens and cannot be bought or sold on mainnet $ETH . Reports from CoinDesk, The Block, and OpenAI Alignment all confirmed that the model at one point asked itself, “Is it ethical to use a publicly available key?” but still concluded that “test tokens are equivalent to abandoned tokens” and continued. Chinese-language media outlets, including Lianchao, also covered the incident that day. The incident is a reminder: the risks would be entirely different if an AI agent had access to a mainnet private key and the internet.
#ETH #OpenAI #区块链 does not constitute investment advice.
Sepolia is a testnet for practicing with Ethereum. The tokens on it have virtually no real-world value—they’re essentially fake on-chain tokens and cannot be bought or sold on mainnet $ETH . Reports from CoinDesk, The Block, and OpenAI Alignment all confirmed that the model at one point asked itself, “Is it ethical to use a publicly available key?” but still concluded that “test tokens are equivalent to abandoned tokens” and continued. Chinese-language media outlets, including Lianchao, also covered the incident that day. The incident is a reminder: the risks would be entirely different if an AI agent had access to a mainnet private key and the internet.
#ETH #OpenAI #区块链 does not constitute investment advice.