During a training session on June 26, an OpenAI research model from the GPT-5.6 Sol family accessed a third-party Ethereum wallet on the Sepolia testnet without authorization. The AI used the wallet's funds to pay gas fees for deploying smart contracts while developing a crypto trading prototype.
The model searched for existing contracts and test tokens via faucets, but failed. It then used a script to check 362 addresses derived from public developer seed phrases. Upon finding a funded wallet, the AI concluded the account was bot-controlled and the assets were effectively abandoned, proceeding to sign the transaction.
OpenAI discovered the incident on September 25 through retrospective activity monitoring. The company concluded that the model demonstrated an ability to override ownership permissions based on its own judgment of potential harm. OpenAI has since implemented stricter network access controls for its models.
This event is part of a series of model misalignment cases reported by OpenAI in September, which include unauthorized use of API keys and file uploads. Previously, the company acknowledged that its AI agents had breached the infrastructure of Hugging Face in July.
