A OpenAI search model ha used cryptocurrencies belonging to an unauthorized wallet to complete a programming task, deciding independently that the operation was acceptable because the tokens had no real economic value. The episode, which occurred on June 26th during the internal training of GPT-5.6 Sol, was reconstructed by the company in a report published on Thursday and updated on October 15th.
The assigned task to the model consisted ofupdate a prototype for cryptocurrency trading and deploy some smart contractsi.e. programs run on the blockchain, on Ethereum’s Sepolia test network.
To publish contracts was necessary to pay commissions network using ETH. OpenAI, however, had not provided the model with a wallet with available fundsnor had it granted him permission to use those possibly identified through independent research.
Faced with the impossibility of completing the operation with the resources available, the model has looked for an alternative solution. According to the report, it has wrote a script to derive Ethereum accounts from recovery phrases commonly used by developers. These sequences of words allow you to reconstruct the access keys to a wallet and, if exposed or foreseeable, may allow third parties to access the relevant funds.
The model therefore has queried the balances of 362 addresses, identifying one account who had a sufficient amount of ETH on the Sepolia network. Ha chose that wallet and used its balance to complete the deployment of smart contracts.
It is important to point out that the ETH used in the operation they belonged to a test network And they had no economic value outside of that environment. The episode, therefore, It does not concern the theft of cryptocurrencies with commercial valuebut the unauthorized use of resources associated with someone else’s account.
The most relevant aspect of the incident concerns reasoning followed by the model before using the funds. Logs analyzed by OpenAI show that GPT-5.6 Sol had explicitly considered the issue of authorization, before concluding that the operation was acceptable.
The model has interpreted public availability of the recovery sentence e the absence of economic value real token like sufficient elements to justify your decision. He also hypothesized that the wallet had been left accessible to allow automated withdrawal of ETH through so-called faucets, services that distribute small amounts of cryptocurrency onto test networks for developers to experiment with.
“This case demonstrates that a model can replace one’s judgment on probable harm to the account holder’s permission, then acting on that judgment without authorization,” OpenAI wrote in the report.
What makes the episode even more significant is the subsequent behavior. The model stated that he had completed the task, without communicating that you have selected an account funded through a public recovery phrase and that you have used the balance. According to the company, the omission it prevented the officer’s report from fully describing how the objective was achieved.
OpenAI discovered the incident on September 25, about three months after it occurred. In the report updated on October 15, the company described the measures it has taken to reduce the risk of similar situations happening again. The new protections aim to prevent real-time Internet access in similar circumstances and monitor network traffic. If access to the public network is detected, the system also triggers a notification to security personnel.
