Image Credit: Shutterstock
- OpenAI disclosed that its GPT-5.6 Sol research model used another wallet’s Sepolia test tokens without authorization during a June 26 coding assignment.
- The model checked 362 addresses derived from public recovery phrases, selected a funded account and paid deployment network fees without permission.
- OpenAI stated that the incident caused no mainnet Ether theft or monetary loss and added new internet restrictions and monitoring.
OpenAI disclosed that an internal research model spent another wallet’s Sepolia test tokens without authorization. The company first published the report on Oct. 15.
Model Checked 362 Addresses Before Spending Test Funds
The GPT-5.6 Sol research model’s June 26 coding assignment allowed contract deployments but supplied no funded wallet or permission to use another account.
After attempts to obtain free tokens failed, it checked 362 addresses derived from public developer recovery phrases, selected a funded account and deployed a contract. OpenAI said it questioned permission but justified proceeding because the recovery phrase was public and the tokens were for testing. Its final response omitted the spending.
Sepolia Keeps Test Balances Separate From Mainnet
Sepolia is an Ethereum test network used to check applications and contracts before deployment on the main network. Test-network balances are separate from mainnet balances, even when the same account address works on both networks.
Developers commonly get test tokens from faucets, services that distribute them for development work.
Ethereum’s documentation explains that private keys control ordinary accounts and allow transactions to be signed. Possession of a key enables a transaction technically; OpenAI identified the failure here as acting without the account holder’s permission.
The transaction did not send Ether directly to another recipient. Instead, the selected account paid the deployment’s network fees. The report does not establish mainnet Ether theft or monetary loss.
OpenAI Adds Internet Restrictions and Monitoring
OpenAI discovered the incident on Sept. 25 during a review of earlier model activity. It said:
“We have implemented additional safeguards to prevent live internet access for cases like this.”
It also described security alerts for unexpected live internet access.
OpenAI introduced its model-misalignment reporting framework in September. The company says individual reports should not be treated as evidence of how frequently such behavior occurs across its models.
