CVE-2016-2384: arbitrary code execution due to a bug in usb-midi linux driver
xairy.github.io
xairy.github.io
There's a narrow problem here, which is the bug in the device driver. That'll be fixed of course, now that it's gotten attention. But fixing this bug will not solve the problem, because there is a truly enormous number of different USB device drivers, most of which are poorly audited, run in kernel space, and correspond to unusual hardware. The real problem is that plugging in a device can cause any obscure device driver to load, without user confirmation. So for example, someone can plug in a cell phone, and the cell phone can claim to be an obscure kind of printer that no one has seen since the 90s, and the system will go ahead and load that device driver.
The solution is to require user confirmation before accepting any USB device that hasn't been seen before. Just a simple popup: You have connected [type of device], Ok/Cancel. The only tricky part is input devices like keyboards and mice, where this could create a chicken-and-egg problem with the confirmation prompt. (And devices pretending to be keyboards can be dangerous; they can type commands). Fortunately, there is a solution, albeit one which will take some programming work: accept keyboard input from keyboards that have either been confirmed by a prompt or have typed the user's login password. (Since some "keyboards" are actually things like barcode scanners).
This is also useful to disable modprobing obscure networking protocols just by an unprivileged process calling socket(SOCK_STREAM, AF_FULLOFBUGS, 0), which was a fruitful source of exploits at least a few years back.
However, for most server machines, if an attacker can plug a malicious USB device into your machine, something has already gone very wrong. (And for a good chunk of servers these days, they don't even have USB because they're VMs.) The primary threat model for this sort of thing seems to be laptops that are briefly left unattended and locked, or "Hey, can I charge my phone off your laptop."
For the unattended-and-locked case, perhaps you could block module autoloading until the user has logged back in. It's roughly similar to the approach Apple takes with Thunderbolt, I believe.
There are a million easier ways to break into a virtual machine running on a physical machine you control than finding an exploit and writing code to emulate malicious hardware. You can just modify memory directly.
However, setting kernel.modprobe to /bin/false is extremely high-impact in reduction of attack surface compared to how much effort it is (one line in /etc/sysctl.d/); you can keep running a full-featured distribution kernel, even one from Fedora or Ubuntu that's simultaneously intended for desktop use.