In a modern operating system the kernel, the core of the OS, in complete control of how much time it allocates to the various user processes it's managing. It can interrupt the execution of a user process through various mechanisms provided by the CPU itself. This is called preempting the process and can be done on a schedule, like executing a user process for a particular number of nanoseconds before automatically interrupting it.
In older operating systems, like DOS and Windows 1.0 through 3.11, macOS 9 and earier, plus many others, they employ a different mode where the user process is responsible for yielding control. If the process doesn't yield there may be little recourse to reassert control of the system. This can lead to crashes or lock-ups, a frequent problem with non-preemptive operating systems of all stripes.
Even then there is often hardware support for things like hardware timers that can trigger a particular chunk of code on a regular basis which can be used to rescue the system from a run-away process. Just because a bit of code is running is no guarantee that it will continue to run indefinitely, without interruption.
A modern CPU is a fantastically complicated piece of equipment. Those with support for things like CPU virtualization can make the single physical CPU behave as if it's a number of virtual CPUs all sharing the same hardware. Each of these virtual CPUs is free to do whatever it wants, including dividing up its time using either a pre-emptive or cooparative model, as well as splitting itself into even more virtual CPUs.
The long and the short of it here is to not assume that the kernel must be actively executing to be in control. It has a number of tools at its disposal to wrest control of the CPU back from any process that might be running.