You could get a Solarwinds type situation where the adversary has the signing keys and ability to publish to the website.
You might also find that the vendor ships a library (like libxz) as a part of their invisible or hidden supply chain, that is able to be compromised.
You might find that one of the people working at the company makes a change to the code to enable remote access by the adversary in a targeted collaboration/attack.
The problem isn't that signing key (although I could delve into the lengths you'd need to go to to keep that secret under these threat models) - the problem is what they sign. A signed end release binary or series of packages isn't going to address the software source code itself having something added, or the dependencies of it being compromised.
Not patching is insane -- you'll let script kiddies in. Patching might not stop the next Stuxnet author, but you'll slow them down _and_ have fewer script kiddies.
A lot of people seem to be focusing on how the band-aid of automatic security updates can be ugly without considering the hemorrhaging that it's actually stemming. Nobody's stepping up with a realistic solution to the problem, which means we're stuck with the band-aids.
[1] https://www.usenix.org/system/files/1401_08-12_mickens.pdf
That's before even addressing mistakes.