AI Coding Agents Fail in Both Directions: Refusing Safe Work and Running Dangerous Commands HackerNoon