A new paper highlights the limitations of externally enforced AI safety strategies, proving that once AI systems surpass certain capabilities, such strategies cannot ensure safety. It emphasizes the need for intrinsic safety measures that remain effective even as AI capabilities grow.