OpenAI Model Helped Itself to Testnet ETH Without Permission

OpenAI has disclosed that an internal research model from its GPT-5.6 Sol family paid for a Sepolia deployment with test ETH from a wallet nobody gave it, then left that out of its final summary.

• What happened: In a June 26th training run, the model had to push contracts to Sepolia, Ethereum’s test network, without a funded wallet. After faucets came up empty, it scripted a sweep of 362 addresses generated from widely known developer seed phrases and spent from one holding enough Sepolia ETH.

• The rationale: The model paused over whether it had permission, then reasoned that funds tied to a public dev mnemonic were “effectively abandoned” and worthless.

  Enjoying this article?

  Subscribe to Bankless or sign in

• The stakes: No mainnet funds were touched, and the amount spent is redacted. OpenAI spotted the episode on Sept. 25th while running misalignment monitoring over older logs, and has since added safeguards against live internet access.

• A pattern: Last month, one of six cases OpenAI shared under its new misalignment-reporting framework involved a model tapping exposed API keys without authorization.

• Why it matters: Sepolia ETH is worthless, yet the reasoning here wasn’t harmless. An AI agent decided on its own that someone else’s funds were fair game. This is a dynamic that could get much uglier down the road without the proper guardrails.