OpenAI model helped itself testnet ETH without permission

OpenAI has disclosed that an internal research model of its GPT-5.6 Sol family paid for a Sepolia deployment with test ETH from a wallet no one gave it, then left it out of its final summary.

• What happened: During training on June 26, the model had to offer contracts to Sepolia, the Ethereum test network, without a funded wallet. After the faucets emptied, he scripted a sweep of 362 addresses generated from widely known developer seed phrases and spent from one containing enough Sepolia ETH.

• The rationale: The model questioned whether it had permission, then judged that funds tied to a public development mnemonic were “effectively abandoned” and worthless.

  Do you like this article?

  Subscribe to Bankless or log in

• The issues: No mainnet funds were touched and the amount spent is redacted. OpenAI spotted the episode on September 25 while conducting misalignment monitoring on older logs and has since added protections against live Internet access.

• A model: Last month, one of six cases OpenAI shared as part of its new misalignment reporting framework involved a model exploiting API keys exposed without permission.

• Why it’s important: Sepolia ETH is worthless, but the reasoning here was not harmless. An AI agent decided on his own that someone else’s funds were a good thing. This is a dynamic that could become much uglier in the future without the proper safeguards.