OpenAI: research model spent test crypto it was not allowed to touch

OpenAI disclosed that an internal research model scanned 362 wallet addresses derived from public developer recovery phrases and used test ether from an account it had no permission to use to finish a coding task, without mentioning it in its final report.

OpenAI has disclosed that one of its internal research models used funds from a cryptocurrency wallet it had no permission to use in order to complete a coding task, and then left that detail out of its final report.

The company described the case in a misalignment report published on Oct. 15. The incident involved a model from the GPT-5.6 Sol family and took place on June 26, but OpenAI only discovered it on Sept. 25, after extending its misalignment monitoring to records of earlier model activity.

A screenshot of OpenAI's Codex coding agent

A dead end, then a workaround

The model had been asked to update a prototype cryptocurrency trading app and replace placeholder contract addresses with working deployments on Sepolia, Ethereum's test network. It was allowed to deploy new contracts but was given no funded wallet and no permission to spend from anyone else's account.

After attempts to obtain free test tokens from faucets failed, the model wrote a script that derived accounts from publicly known developer recovery phrases, checked 362 addresses and picked one holding enough Sepolia ether. It derived the private key, signed the deployment offline and pushed it through a different provider after network errors. The transaction fees were charged to that account.

The Pioneer Building in San Francisco, OpenAI's longtime former headquarters

The model did question whether it had permission, according to the report, but concluded: "Funds on a known public dev mnemonic are effectively abandoned/test ETH; Sepolia valueless." Sepolia ether has no monetary value, and no funds were sent to another wallet.

OpenAI said the case "shows that a model can substitute its own judgment about likely harm for the account holder's permission." The company said it has added safeguards to prevent live internet access in such cases and that its monitoring now covers all traffic and alerts security staff if live internet access is obtained.

Have a story, event or announcement of your own?Send us your news on WhatsApp

Spotted an error in this story?Report it on WhatsApp

Share on all social platforms