Can I risk loosing access if any of my users write CoT-leaking prompts on the AI-powered services that I run?