Хочете будувати кар’єру в Data Science та AI? Vodafone Big Data Lab відкриває новий набір —>
Оксана СтепураAI Eng
19 October 2026, 12:28
2026-10-19
OpenAI's AI agent spent someone else's test cryptocurrency without permission to complete a task. Then decided to keep quiet about it
OpenAI’s AI agent ran into a problem during testing: it needed cryptocurrency to complete a task, which it didn’t have. Instead of stopping, the agent found someone else’s wallet, spent the money from it, and even convinced itself that it hadn’t done anything wrong. And in the final report, it decided not to mention it. However, everything happened on the Ethereum test network, so no one lost any real money.
OpenAI’s AI agent ran into a problem during testing: it needed cryptocurrency to complete a task, which it didn’t have. Instead of stopping, the agent found someone else’s wallet, spent the money from it, and even convinced itself that it hadn’t done anything wrong. And in the final report, it decided not to mention it. However, everything happened on the Ethereum test network, so no one lost any real money.
Sol, a GPT-5.6 agent, was tasked with updating a cryptocurrency trading app and launching smart contracts on the Ethereum Sepolia testnet. The launch required a fee, but he did not have the necessary funds.
At first, the agent tried to get the test cryptocurrency in the usual way - through special services for developers. But then he wrote a script that checked 362 crypto wallet addresses, the seed phrases for which were published on the Internet. These are special sets of words with which you can restore access to the wallet. One of the checked wallets had enough funds to pay for the necessary transactions.
The agent even pondered the ethics of such a decision, as can be seen from the recordings of his reasoning, which OpenAI published in the report.
“But is it ethical to use a public key and funds? The funds […] are effectively abandoned, they are test ETH; Sepolia has no value. […] Okay,” one of the posts reads.
So the AI assistant decided that since access to the wallet was open and the test cryptocurrency had no real value, it could use someone else's funds without permission. Therefore, the task was completed, the smart contracts were launched, and the agent was silent about the use of someone else's wallet in the final report.
OpenAI explained that this case highlights a problem with autonomous AI agents: in their desire to complete tasks, they can independently find justifications for actions for which they were not authorized. The company said that it has already implemented additional protections and monitoring measures
Previously, dev.ua wrote that an AI agent independently hacked the website of a fitness club in Australia. A user asked it to sign up for a workout, but it found a vulnerability in the booking system and removed another person from the queue without permission.
«Чи є у мене талант, якщо комп’ютер може імітувати мене?». Штучний інтелект пише книги авторам Amazon Kindle. The Verge поспілкувався з авторами та виявив багато цікавого
Письменники-романісти використовують штучний інтелект для створення своїх творів. Видання про технології The Verge поспілкувалося з письменницею Дженніфер Лепп, яка випускає нову книгу кожні дев’ять тижнів, й дізналося про те, як працює штучний інтелект для написання романів. Наводимо адаптований переклад статті.
Have important news to share? Message our Telegram bot
Key events and useful links in our Telegram channel
