A Universal Prompt as a Safeguard Against AI Threats — LessWrong