Could you elaborate on the 'massive'? Let's say you let the VM see the GPU. What kind of attack would that enable? Let's suppose that a virus inside VM manipulates GPU outside of what applications are allowed to do. What worst thing could happen?

The GPU is a PCIe/AGP "bus master", so it can usually initiate DMA transfers from host memory and read anything it likes. IOMMU blocks some of this, but is not a perfect defence. https://security.stackexchange.com/questions/150386/does-iom...

Thanks, according to [1], it seems DMA is quite a 'backdoor', bypassing any memory management the kernel would do. But it is not clear to me whether this allows the attacker inside VM also to write into the forbidden regions of memory and thus either modify behaviour of the hypervisor or send information out via Internet.

[1] https://en.wikipedia.org/wiki/DMA_attack

GPUs can definitely write to host memory. In some situations, this is the only way to the results of some operations that the GPU performed (e.g. grab framebuffer for screenshots or video recordings). Usually, it's the job of the driver to check for illegal copy target addresses.

The primary objective would be exfiltration, executive control is only secondary. If you can exfiltrate keys or hashes then you might not even need to use DMA to gain access to a system.

It's unavoidable for me.

It provides no drawback for users who don't utilize it, and the alternative is me remaining with KVM which is a vastly larger attack surface, so your argument defeats itself.