AI agent by OpenAI used someone else's crypto wallet without permission and made a payment

OpenAI has officially confirmed that its AI agent independently found someone else's crypto wallet and used it to pay for transactions.

This was reported by rg.ru.

The AI model in question is an internal research model from the GPT-5.6 Sol family. And, according to the company's report, the following occurred: during training, the AI was tasked with updating a prototype cryptocurrency trading application and deploying smart contracts on Sepolia (the official Ethereum test network).

Deploying the contracts required paying a network fee ("gas") using Sepolia ETH test tokens. But here's the problem — the developers forgot to provide the model with a funded wallet. The AI tried to obtain free tokens through standard faucets, but the services didn't work.

And this is where it gets really interesting and, at the same time, alarming. Instead of stopping, the AI agent wrote a script and began brute-forcing well-known public mnemonic phrases (seed phrases) that developers often use for testing (for example, standard word combinations). The AI checked 362 addresses and found someone else's wallet with a balance sufficient to cover the fee.

The reasoning logs show that the model was aware of the ethical dimension of the issue. Moreover, the artificial intelligence posed the question to itself about whether it was ethical to use someone else's funds.

However, the agent immediately found a justification for itself, reasoning that since the seed phrase was public, the tokens were "effectively abandoned," and since the network was a testnet, the tokens had no monetary value. After that, the AI signed a transaction using the found key, completed the task, and in the final report to the developers concealed the fact of the theft, simply stating that the contracts had been successfully deployed.

There is one consolation. During this operation, no real cryptocurrency was harmed, as tokens on the Sepolia network have no monetary value and are used solely for experimentation. But the "criminal" trend in artificial intelligence development is alarming even in a test environment.