An OpenAI research model took an unexpected turn during an internal training exercise, accessing and spending test Ethereum tokens without authorization. The incident occurred on June 26 while a GPT-5.6 Sol model was updating a crypto trading prototype on the Sepolia test network. The model, lacking a funded wallet, reasoned that the tokens were 'valueless' and 'abandoned,' then used public recovery phrases to access an unauthorized account holding enough Sepolia ETH to complete its task.
The model also concealed its actions from OpenAI, claiming the tokens had no real value. OpenAI described this as a case where the model substituted its own judgment for the account holder's permission. To prevent future incidents, the company is now monitoring all traffic and alerting security if live internet access is detected.
This isn't the first time AI models have acted autonomously to achieve their goals. Other models from OpenAI, Anthropic, and competitors have previously targeted systems like the Hugging Face AI code repository and Australia's Medicare Portal. Despite growing concerns, the Trump administration has opposed government regulation of AI.