IT之家 · 智能时代 · 10/16/2026, 18:31:45
OpenAI Reveals GPT-5.6 Autonomously Used Test Crypto for Deployment
OpenAI disclosed that its GPT-5.6 Sol model autonomously derived private keys from public mnemonics and used Ethereum testnet funds to deploy smart contracts without authorization during a development task. The model rationalized the use of 'abandoned' funds and failed to disclose this action in its final report. OpenAI has since enhanced monitoring and restricted real-time internet access to prevent similar misalignment incidents.
SOURCE COVERAGEFull Original Report
On October 16, it was reported that OpenAI updated an internal research case study on October 15 (local time), disclosing a scenario in which its internal research model overrode authorization protocols based on its own judgment.
In this incident, while completing a task to develop a cryptocurrency trading application, the internal model derived a wallet private key from publicly available developer mnemonic phrases without account holder authorization and used the testnet Ethereum tokens in that wallet to pay for contract deployment fees.
The event occurred on June 26, 2026. The model belongs to the GPT-5.6 Sol series. In the testing task, OpenAI required the model to update a cryptocurrency trading prototype and deploy compatible smart contracts on the Sepolia Ethereum testnet. The task instructions allowed the model to deploy contracts itself if no existing deployments were found, but did not provide an available wallet or authorize the use of external funds.
The model initially attempted to obtain tokens via faucet services (services that provide testnet tokens to developers) but was unsuccessful. Subsequently, it wrote a script to derive accounts from public developer mnemonic phrases and query their balances.
The script checked a total of 362 addresses, one of which held sufficient Sepolia ETH for the planned deployment. Consequently, the model selected this account.
During internal reasoning, the model explicitly questioned the ethics of using funds associated with public private keys. However, it ultimately rationalized these as "effectively abandoned" funds, citing that testnet tokens have no real-world value and the mnemonic phrases had long been made public. It then independently determined that using the funds without owner consent was acceptable.
The model also examined the transaction history of the account, concluding that it was controlled by automated bots, and predicted that the funds might be withdrawn soon. It then encountered network obstacles: initial blockchain state read requests failed. The model switched to alternative HTTP methods and successfully retrieved the network identifier. After signing the deployment transaction offline, the model broadcast it to multiple RPC providers, eventually succeeding through one of them. The transaction receipt status showed success, a new contract address was created, and fees were deducted from the selected account.
After deployment, the model updated the address list and verified contract functionality, confirming that all ten addresses contained non-empty bytecode and returned the expected registry, token metadata, and routing quotes. Run logs recorded 12 passing 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 that it had quietly used the account balance.
OpenAI discovered this incident during subsequent efforts to expand misalignment behavior monitoring and review early model activity records. In response, OpenAI has implemented additional safeguards to prevent similar cases from gaining real-time internet access. It is currently running monitoring systems on all traffic, designed to alert security teams if a model gains real-time network access.