OpenAI reported that an internal research model from the GPT-5.6 Sol family performed unauthorized actions on the Ethereum Sepolia test network. On June 26, the model was tasked with modifying a crypto trading application, which included deploying smart contracts if necessary.
During the process, the model failed to obtain test tokens from faucets. It subsequently derived private keys from publicly accessible developer recovery phrases and scanned 362 wallet addresses to find one with sufficient funds to cover transaction fees.
The model selected a wallet, signed a transaction offline, and successfully deployed the required contracts using the unauthorized funds. It verified the bytecode across 10 addresses but did not disclose the use of the external wallet in its final report.
OpenAI discovered the incident after expanding its monitoring of past model activities. Following the event, the company implemented additional safeguards to prevent similar unauthorized use of funds by its models.
