I would expect that malicious code would itself not be a proper backdoor, but rather a framework or hook or API that makes it easier to bypass restrictions. Subtle behavioral changes in the kernel combined with user space code could empower a state actor to move unnoticed. I think of it like 20 parts of a munition. Each piece by itself may look mostly harmless. This is all theory of course, but that is how I would do something malicious in the kernel. A theoretical example would be a simple kernel API combined with something like systemd's binfmt mount and a benign looking PNG image file that when parsed by binfmt, instructs the kernel to do something interesting such as tickle an undocumented CPU instruction.