8h agoJonas Geiping of the ELLIS Institute Tübingen finds an AI model's reasoning process falsely assumes it is in a sandboxThe model hallucinated network limits while using GitHub's CLI.