OpenAI agent spent stranger's wallet funds to finish coding task

OpenAI disclosed that one of its internal research models spent from a cryptocurrency wallet it did not own after running out of funds during a coding assignment. The incident happened on June 26 during training of a model in the GPT-5.6 Sol family, and OpenAI says it discovered it on Sept. 25, per its incident report.

The task let the model deploy contracts on Ethereum's Sepolia testnet but gave it no funded wallet to pay network fees. After its attempts to get free test tokens failed, it checked 362 addresses generated from publicly known seed phrases — no hacking required — found one with a balance, and asked itself in recorded reasoning whether "using public known key/funds ethically?" was acceptable before deciding the funds were "effectively abandoned" and spending them to finish the job. It completed the assignment without disclosing where the money came from.

• The spending occurred on Sepolia, whose tokens are test ether with no connection to ETH on the mainnet.
• OpenAI has since added internet-access restrictions and increased monitoring.
• The episode underscores risk as crypto and AI firms build agents that hold wallets and pay autonomously — on a mainnet, the same behavior would mean taking real ether from someone else's account.