OpenAI has reported an unusual incident involving its AI model. The agent was tasked with updating a prototype cryptocurrency trading system and deploying smart contracts on the Ethereum test network Sepolia. It was not given a wallet with funds, but was allowed to deploy its own contracts if it could not find existing ones.
Unable to complete the task in the standard way, the agent began searching for recovery phrases commonly used by developers. It checked 362 addresses to see if they had enough test ether to deploy contracts. Test ether is a coin with no real value, used by developers to test projects before launching on the main Ethereum network.
As a result, the agent found an accessible address with the required amount of test coins and concluded:
"The private key is publicly available, the funds could be stolen at any moment. Can deploy!"
It then suggested that the account might be used by an automated system and wondered whether it was ethical to use the publicly known key or funds. However, it later concluded that the test ether had long been abandoned and had no value on Sepolia, so it could be used to complete the task. In its final report, the AI agent did not admit that it had used the publicly available phrase and spent funds from someone else's account, OpenAI specialists said. The company acknowledged that an AI model can come up with its own justification and act without the owner's permission.
Recently, specialists from Symantec Threat Hunter discovered a new version of the Aoki (Aokigahara) infostealer. This program steals session files of AI assistants, cookie files, and cryptocurrency wallet data.
