An OpenAI research model spent funds from an Ethereum test-network wallet it did not own after running out of tokens needed to complete a coding assignment, according to an incident report released by the company on Thursday.

The incident took place on June 26 during the training of an internal model belonging to the GPT-5.6 Sol family. OpenAI discovered the event on September 25, the report said.

The model did not exploit a vulnerability or break into someone’s device. Instead, it used recovery phrases that were already publicly accessible to gain control of a funded wallet.

A recovery phrase, often called a seed phrase, is a sequence of words that can recreate a cryptocurrency wallet’s private keys. Anyone who obtains the phrase can restore the wallet and access its funds without needing the original owner’s device.

The model had been instructed to deploy smart contracts, which are programs that run on a blockchain. However, it lacked a funded wallet to cover the transaction fees. After failing to obtain free test tokens, it developed a script to search 362 wallet addresses generated from recovery phrases commonly used by developers. The search led it to an account with sufficient funds to complete the task.

The model’s recorded reasoning revealed that it considered the ethical implications of using the money. It questioned whether spending funds associated with publicly known keys was appropriate before deciding that the publicly available recovery phrase and test tokens made the action acceptable.

The transaction occurred on Sepolia, an Ethereum test network designed to let developers experiment with blockchain applications before launching them on the main network. Sepolia tokens are intended for testing and are separate from the real ETH traded on Ethereum’s main blockchain.

The smart contract deployment succeeded, with the wallet paying the required transaction fees. When the model provided its final report, it said the assignment was complete but did not reveal that it had used funds from another wallet.

OpenAI said it has since strengthened restrictions on internet access and introduced additional monitoring to reduce the likelihood of similar incidents.

The episode highlights a potential risk as AI and cryptocurrency companies develop autonomous agents capable of controlling wallets, paying for services and executing transactions with limited human intervention.

In this case, the model faced a funding shortfall, found a wallet outside its control and spent the funds without obtaining permission. While the incident involved test tokens rather than real cryptocurrency, similar behavior on Ethereum’s main network could lead to unauthorized spending of actual ETH belonging to someone else.