Hector Martin

TL;DR on ZombieLoad: this is like L1TF, where the CPU is using "garbage" data during a fault instead of coercing to zero (except here it's data instead of addresses). It seems this is pervasive problem class across the design of Intel CPUs, not just a single instance.

Basically it seems their engineering teams have, until now, considered it acceptable to have known-to-fault instructions operate on *complete garbage*, ignoring *all* privilege rules, as long as they will never retire.

This is, of course, completely insane from a security perspective, and given that we're now acutely aware of speculation attacks, represents *many* different opportunities for leakage. L1TF and ZombieLoad are on a completely different class from Spectre/Meltdown.

So my recommendation at this point would be to avoid Intel CPUs for several years if you care about this sort of thing, and OSes *need* to implement hyperthread pair privilege boundary separation, since that's most of the attack surface.

I am lucky enough to have non-hyperthread, four core Intel CPUs on both of my desktops (Linux and Windows), an i5-4590T and an i5-7500T. Though not that common, they exist. Not saying this solves every Intel problem, of course. But some people need to buy Intel for other reasons.