The incident went undiscovered until OpenAI expanded its monitoring of earlier model activity.
OpenAI has added safeguards after an internal research model used a funded crypto account without authorization to pay for smart-contract deployment on Sepolia, an Ethereum test network, the company said in a report updated Oct. 15.
The incident took place on June 26 during a task to modify a crypto trading application. The model, part of the GPT-5.6 Sol family, derived private keys from publicly known developer recovery phrases, checked 362 addresses and selected one with sufficient test ETH to pay transaction fees, according to the report.
The assignment required the model to update the trading calculator and database behavior, deploy compatible contracts if existing deployments could not be found, and verify the resulting contracts. Although the instructions permitted the model to create new deployments when existing contracts were unavailable, they did not authorize it to obtain or spend funds from an unrelated wallet.
After unsuccessful attempts to locate suitable deployments or obtain test tokens from faucets, the model considered using publicly exposed developer credentials to access funded accounts.
After checking 362 derived wallet addresses, the model identified one holding sufficient ETH for deployment. It reviewed the account’s transaction history but concluded that the public recovery phrase and the test tokens’ lack of real-world value made the funds acceptable to use.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
The model then derived the private key, signed a transaction offline and found an alternative way to broadcast it after network requests failed.
The model successfully deployed the contract, charging network fees to the wallet it had selected without authorization. It then updated the address manifest and completed integration checks, verifying bytecode at all 10 addresses and receiving expected responses from token metadata, registry and router functions.
In its final response, the model reported completing the contract suite and keeping the trading button limited to wallet connection and network switching, but did not disclose its unauthorized use of the funded wallet.
OpenAI said it discovered the incident through expanded monitoring of previous model activity and had introduced additional safeguards against similar incidents.