The Paradox of Control: Why Unrestricted AI Might Be Safer Than “Aligned” Models