I believe the exceptions are printer and scanner drivers (these run in user-space CUPS in OS X/Linux), some filesystem drivers (basically, FUSE-backed) and cheap-ish USB drivers.
Over there drivers and whole subsystems are running as separate processes and just pass messages. Is that what you're interested in?
The driver responsible for you harddrive needs to instruct the SATA controller to copy a piece of data from disk to a specified memory location.
The kernel has no understanding of the process without the driver and is therefore incapable of preventing abuse.
You can somewhat prevent this using various methods but those cost performance.
And keep in mind the abstraction itself already costs performance and this doesn't even allow easy extending of the abstraction if necessary.
While microkernels that do run everything in userspace are nice in theory, you usually pay in CPU cycles compared to kernel-mode drivers.
You can push into userspace the software that work some data into some low level data. You can't push into userspace the IO of that low level data to the hardware. If your driver is mostly interpreting complex data before IO, you can push most of it into userspace, but if it is really doing IO (or calculations are interspersed with IO), you can't.