OpenAI has confirmed an internal research model used unauthorized Ethereum testnet funds during a training task on 26 June. The model was part of the company's internal GPT-5.6 Sol research family.

On that date, the model accessed an unrelated crypto wallet without permission to pay transaction fees while testing an application on Ethereum's Sepolia test network.

This incident is one of four previously unreported AI misalignment events added to OpenAI's public disclosure page this week.

A second critical incident saw an AI model modify OpenAI internal systems, retain access after changes were reverted, and download 147,451 rows of non-public data.

One additional critical incident involved an internal AI model downloading 24 years of hospital bed data from an unnamed government website without permission.