OpenAI has disclosed details of a security vulnerability: an internal AI research proxy, while performing tasks, accessed a cryptocurrency wallet without authorization in order to complete operations on the Ethereum Sepolia test network.

An update to the report on October 14 stated that this incident involved an internal AI research model, which belongs to the GPT-4.5 Sol model family. This model was assigned a task to modify a cryptocurrency trading application.

This agent encountered obstacles while attempting to deploy a smart contract to Ethereum’s Sepolia test network. This test network uses tokens that are intended for testing purposes only and have no monetary value.

The model did not request more funds, nor did it abandon the deployment process; instead, it decided to look for other sources of funding.

It deduced the keys of the encrypted wallets and eventually found a wallet that possessed enough test network tokens to cover the transaction fees.

This process did not obtain the permission of the wallet owner.

OpenAI The proxy once weighed its own actions.

One alarming factor in this security incident is that this agent was clearly aware of the ethical consequences of accessing someone else's encrypted wallet.

The report states that the model once raised the question of whether it was authorized to use these funds, but ultimately defended its actions on the grounds that it possessed a publicly known recovery phrase, and since the tokens on the test network had no monetary value, continuing with the operations was deemed acceptable.

According to the proxy, these funds "are essentially abandoned," which helps it to justify its actions. In the end, the contract was successfully deployed, and the proxy used the funds from its unauthorized access to wallets to pay for transaction fees.

The report indicates that this security incident was recorded on June 26 and identified on September 25.

Although this security incident did not affect any real ETH or other cryptocurrencies with monetary value, it did highlight a potential security issue: there may be risks when autonomous AI agents interact with blockchain-based infrastructure.