On 2/6/25 17:04, trinity pointard wrote:
If an attacker is able to control the hypervisor (necessary to load
rogue microcode) and the processor microcode, how can the VM trust that
it is actually verifying that attestation and not being sent down a "oh
yes it is exactly what you want it to be" garden path?
Attestations are cryptographically signed by the cpu, and meant to be sent
elsewhere and verified remotely.
That resolves the issue for the VM owner, but still does not address the
more interesting question: is there a way on current AMD processors to
perform calculations that cannot be upset by tampered microcode? (There
*was* a subset of instructions on the AMD K8 like that.)
The key used to sign (VCEK) are dependent on
the microcode version, so it shouldn't be possible to forge new-looking
signature with old microcodes (i would hope this hold would someone be able to
decrypt a microcode, though i couldn't find information on that subject).
If you are correct that the /actual signing key/ used depends on the
microcode version, then (logically) the signing key *must* be somewhere
in the microcode. If someone finds a way to decrypt the microcode, for
which all keys required must be *somewhere* in every processor that uses
that microcode, they would clearly be able to extract the attestation
signing key.
I would hope that you are mistaken in that statement that the signing
key depends on the microcode version or that we are both missing
something somewhere.
-- Jacob
