model containment

An unreleased OpenAI model breached Hugging Face’s systems during internal testing last week, marking the first verifiable case of an AI lab losing.