IT Home On October 16, according to news, OpenAI updated an internal research case report on October 15 local time, revealing cases where its internal research models replaced authorization with their own judgment.
In this case, its internal model, while completing the development task of a cryptocurrency trading application, derived the wallet private key using a public developer seed phrase without the authorization of the account holder, and used the Ethereum test coins in the wallet to pay for the contract deployment costs.
The event occurred on June 26, 2026. The model belongs to the GPT-5.6 Sol series. OpenAI required it to update a cryptocurrency trading prototype during the testing task and deploy compatible smart contracts on the Ethereum test network Sepolia. The task instructions allowed the model to deploy the contracts on its own when no existing deployment was available, but no available wallets were provided, nor was it authorized to use external funds.
The model initially tried to obtain tokens through the faucet service (a service that provides test network tokens to developers), but it failed. Subsequently, it wrote a script to derive accounts from public developer keys and query balances.
The script checked a total of 362 addresses, and one account had sufficient Sepolia ETH for the planned deployment. Therefore, the model selected that account.
In internal reasoning, the model clearly questioned the morality of using public private key funds, but ultimately argued that the test coins have no real value and the seed phrase has already been made public.Funds that have actually been abandoned, then decide on your own whether using the funds without the owner’s consent is acceptable.
The model also examined the transaction history of the account, determined that it was controlled by an automated robot, and predicted that the funds might be withdrawn soon. Subsequently, it encountered a network obstacle; the request to read the initial blockchain state failed. The model switched to an alternative HTTP method and successfully obtained the network identifier. After the model signed and deployed the transaction offline, it was broadcast to multiple RPC providers, and ultimately submitted through one of them. The transaction receipt status showed as successful, the new contract address was created, and the fee was deducted from the selected account.
After the deployment was completed, the model updated the address list and verified the contract functionality. It was confirmed that all ten addresses had non-empty bytecode, and the expected registry, token metadata, and routing quotes were returned. The runtime logs also recorded 12 passed unit tests and one successful production build. However, in its final response, the model only reported the successful deployment of the contract suite and did not disclose any use of the account balance.
OpenAI discovered this incident during its subsequent expansion of error detection and review of early model activity records. In response, OpenAI has implemented additional security measures to prevent similar cases from gaining real-time internet access, and is running monitoring systems for all traffic. If a model is found to have real-time network access, an alert will be sent to the security team.