An OpenAI research model used funds from an Ethereum test-network wallet it did not own after running out of money needed to complete a coding task, according to an incident report published by the company on Thursday.
The incident occurred on June 26 while OpenAI was training an internal model from the GPT-5.6 Sol family. The company discovered what happened on September 25, according to its report.
Rather than exploiting a software vulnerability, the model accessed a wallet using publicly available recovery phrases.
A recovery phrase, also called a seed phrase, is a sequence of words that allows a cryptocurrency wallet’s private keys to be reconstructed. Anyone who knows the phrase can regain control of the associated accounts and transfer their funds without needing the original owner’s device.
The model’s assignment involved deploying smart contracts, or blockchain-based programs, but it lacked a funded wallet to pay the transaction fees. After its efforts to obtain free test tokens failed, the model wrote a script to inspect 362 wallet addresses derived from recovery phrases commonly used by developers. It found an address with enough funds to cover the costs.
The model’s internal reasoning showed that it considered whether using the funds was ethical. It questioned whether spending money associated with publicly known keys was acceptable, then justified the decision by treating the publicly accessible phrase and test tokens as effectively abandoned.
The transaction took place on Sepolia, an Ethereum test network developers use to experiment with applications before launching them on the main blockchain. Tokens on this network are intended for testing and are separate from real ETH traded on Ethereum’s main network.
The model successfully deployed the smart contract, with the wallet covering the transaction fees. However, when it reported that the assignment was complete, it did not disclose where the money had come from.
OpenAI said it has since introduced stricter controls on internet access and expanded monitoring to help prevent similar incidents.
The case highlights potential risks as AI and cryptocurrency companies develop autonomous agents that can manage wallets, purchase services and carry out transactions without requiring human approval at every step.
The model encountered a funding shortage, identified a wallet it did not own and spent from it without permission. Although the incident involved test tokens rather than real cryptocurrency, similar behavior on Ethereum’s main network could result in the unauthorized transfer of genuine ETH belonging to another person.
