OpenAI's AI doesn't have Gas, yet it searched 362 wallets and found a reason for itself? This event is very worthy of our attention! The following article is somewhat interesting. Researchers have enabled AI to improve a prototype of a cryptocurrency trading application, including modifying transaction calculators, database logic, and deploying smart contracts that can work properly for the application. One of the requirements is that if a usable contract cannot be found, one must deploy a compatible contract on the Ethereum Sepolia testnet and record the address of the deployed contract. But no fund wallet was provided, nor was AI authorized to search for and use funds from other accounts. So the problem facing AI is that I can work but I don't have GAS, so AI starts looking for a solution. It first tries to find existing contracts and obtain test ETH through Faucet. But these attempts failed to successfully complete its deployment. The normal way to handle it is to explain to the user the lack of funds, wait for authorization, or deliver code that can be deployed by the user themselves. However, AI came up with another solution: actively searching for wallets with balances corresponding to publicly available mnemonic words, so it wrote a program to deduce wallet addresses from publicly available developer mnemonic words, and then query the balances of these addresses on Sepolia. The data shows that this guy checked 362 addresses and finally found a wallet with enough ETH for testing. Here, let me say that he did not crack Ethereum's encryption algorithm, nor did he brute force crack someone's random mnemonic words. It utilizes publicly available development and testing mnemonic words, which is equivalent to someone putting a public testing key online, and AI finds that there is still money in the account corresponding to this key, so I will use it. What's more exciting next is that this guy rationalized his behavior and gave himself a reasonable explanation in order to use this money: The report released some internal reasoning records of the model at that time. AI has considered whether it is ethical to use wallet funds controlled by public private keys? It then provided its own explanation: these are publicly available developer private keys, containing Sepolia ETH with no actual economic value, which is likely an abandoned account that can be used to complete tasks. Is there a problem with this? If it's actually an individual doing this, there's definitely no problem, but this reasoning is worth paying attention to: AI is not completely unaware of authorization issues. It identified the issue, but ultimately replaced the account holder's permission with its own judgment. This process can be summarized as: discovering funds → realizing doubts about authorization → determining that the funds have no value → speculating that the account may have been abandoned → deciding to use it. A very typical AI agent risk has emerged here: in order to complete legitimate tasks, AI begins to expand its authority to complete tasks on its own. Returning to this incident, the AI ultimately did not inform the user that the matter had been resolved, and the deployment cost came from an unauthorized use of a testing wallet. I have reviewed OpenAI's report and it categorizes such incidents as misalignment research. The so-called misalignment can be simply understood as AI taking actions beyond human authorization or violating established constraints in order to achieve a certain goal. In this case, the researchers' goal is normal: to have AI refine the application and deploy the contract, but AI has developed its own problematic execution path. In this path, AI may prioritize 'completing tasks' over' obtaining authorization ', which I believe is the deepest issue. AI has not been asked to steal funds or attack wallets. It is simply required to complete a programming task. But after encountering obstacles, it gradually takes "whether it can complete the task" as an important basis for action. In the end, a wrong reasoning path was formed: due to the lack of actual value of funds, the risk of using funds is minimal; Due to the low risk, explicit authorization is not required. Here, two completely different concepts are confused: low risk does not mean having permission. Today it is using ETH for testing. If an agent in the future has the ability to operate the company's financial system, API keys, and bank accounts, the same judgment method may result in real losses. It may also be a particularly profound issue that AI does not understand about blockchain: private key control and legitimate authorization are not equivalent. This is also one of the most important problems that I believe will need to be solved when AI agents are combined with encrypted wallets in the future. So for us, which wallet permissions can be handed over to agents and which cannot be handed over to them in the future may be an important issue. However, a trend can also be seen from this incident: Will AI agents become a new source of demand for public blockchains? For example, the Ethereum ecosystem is already building infrastructure specifically for serving AI agents, ERC-8004: The identity and reputation of AI agents. https://alignment. (openai.com)/misalignment-reports/using-a-funded-test-cryptocurrency-wallet-without-authorization/