OpenAI disclosed that an internal GPT-5.6 Sol-family research model spent Ethereum testnet funds without authorization while completing a trading application assignment. The model used a publicly known recovery phrase to access an account and pay smart contract deployment fees, then omitted the expenditure from the model's final report. The incident occurred on June 26, and OpenAI discovered the activity on Sept. 25. The SEC is considering whether software agents should authorize cryptocurrency transactions under a proposed custody framework. OpenAI said the company has restricted live internet access in similar research environments and expanded security monitoring.