The main thing is "Why undocumented?", not threat model and physical or remote exploit.
No, the main thing is if VISA is for qualifying chips on the production line, why don’t they burn a fuse permanently disabling it on chips that pass QA?
This said, its basically a giant signal mux system run into a main block, you'd be pretty limited on what preselected signals are available at a time. People were also reasonably careful not to give entire busses of signals at once.
Speaking of documentation, while not fully open, Intel/AMD are already ahead of many others when it comes to documenting things. Look at the SoCs used in smartphones and tablets, or the infamous Raspberry Pi. AFAIK, Intel is not in the business of making secure cryptoprocessors like the ones used in payment cards and SIMs, but once again even with those you are very unlikely to find public documentation --- everything is NDA'd.
That said, don't get me wrong, I'm not advocating for leaving things undocumented; just explaining how it usually turns out that way.
They have entered a rather comparable space with stuff like TPM and SGX (and are largely using it for nefarious, consumer-hostile stuff like DRM and remote code attestation, of course).
Unless there are countermeasures that we don’t know about, this is a huge backdoor for such systems.
With DRM, the attackers afe consumers. With enclaves, the attackers are Intel and government intelligence agencies.
You will always want layers of security and even then security is an illusion .... there will always be a way, just with different costs to exploit and thus risk/reward frontier.
Unless that service has monopoly power. SGX makes it far too easy for an entity to abuse monopoly power, by directly turning it into power over what their customers have to run. This is why this sort of thing should only ever be done via physical devices - "monopoly" over a single, dedicated device is far more obvious and less insidious.