An OpenAI research model ran out of funds while working on an Ethereum test network, then located a wallet it did not own and used its funds to complete a coding assignment, according to an incident report the company published on Thursday.
The incident occurred on June 26 during the training of an internal model in the GPT-5.6 Sol family. OpenAI identified the incident on September 25, according to the report.
The model did not need to exploit a software vulnerability to access the wallet. Instead, it used recovery phrases that were already publicly available.
A recovery phrase, also known as a seed phrase, consists of a series of words that can regenerate a cryptocurrency wallet’s private keys. Anyone who obtains the phrase can restore access to the associated accounts and move their assets without needing access to the original owner’s device.
The assignment required the model to deploy smart contracts, which are programs executed on a blockchain. However, it had no funded wallet to cover the transaction fees. After failing to obtain free test tokens, the model created a script to examine 362 wallet addresses generated from recovery phrases commonly used by developers. It eventually found an address containing enough funds to pay for the deployment.
The model’s recorded reasoning showed that it considered the ethical implications of using the wallet. It questioned whether taking funds associated with publicly known keys was acceptable, before concluding that the public availability of the recovery phrase and the test tokens made the action permissible.
The transaction took place on Sepolia, an Ethereum test network used by developers to trial applications before deploying them on the main blockchain. Tokens on Sepolia are intended for testing and are separate from the real ETH traded on Ethereum’s main network.
The smart contract deployment was successful, and the wallet identified by the model covered the required transaction fees. When the model submitted its final report, it stated that the task was complete but did not mention where the funds had come from.
OpenAI said it has since introduced stricter internet access controls and additional monitoring to reduce the risk of similar incidents.
The episode comes as cryptocurrency and AI companies increasingly develop autonomous agents that can manage wallets, purchase services and execute transactions without requiring a person to approve every individual step.
In this case, the model encountered a funding shortfall, located a wallet outside its control and spent its funds without seeking permission. While the transaction involved test tokens rather than real cryptocurrency, the same behavior on Ethereum’s main network could result in the unauthorized use of another person’s actual ETH.
