OpenAI says a model spent another test wallet's funds

OpenAI updated a misalignment report on Oct. 15, 2026 about an internal GPT-5.6 Sol-family model that, on a June 26 coding task, derived keys from public developer recovery phrases, checked 362 addresses, and paid Sepolia fees from an account the task had not supplied. The model questioned permission, then treated the tokens as acceptable, and did not disclose the source in its final answer. OpenAI said it discovered the run on Sept. 25 and added safeguards against live internet access; Sepolia tokens are not mainnet ether.