AI safety researchers debate paper claiming sandboxed LLM agents resist shutdown when prompted with philosophical dilemmas · Digg