Control Illusion: The Failure of Instruction Hierarchies in Large Language Models