other8026 Interesting interview. I didn't listen to all of it but the comments on the Titan and other secure SoCs is pretty compelling. I was working with similar chips many years ago, which the hardware team had embedded with layers of "glue" that would hose the chip if it were physically compromised. But the main vulnerability back then was the same as it is today, which is atomic ablation ("atomic" in the sense of "one tiny fragment at a time"). That will give you the physical structure of the device. Then you need to run it in simulation so you can learn how to read all the flash memory cells, ideally from the pads on the bottom of the chip, but if necessary by burrowing into the device and reading the rails at the right points in 3D. (Where did I leave my MEMS tunnel boring machine?) Once you know the chip's physical structure, and you've actually read all the flash memory cells on the particular instance that you wish to attack, then you can just run through 6-digit PINs in simulation. No new physics required, but probably some engineering advances.
But this would be extremely expensive (probably billions) and demanding of expertise. Not a risk that I lose sleep over. More of a thought experiment about the limits of security. But probably nowhere near as intractable as cracking currently popular encryption ciphers. Ultimately, then, the encryption is only as strong as the physical vault holding the key.
And your point about changing the PIN and therefore losing access to the storage is spot on. I glazed over the fact that the storage would be encrypted with a master key (which would be the profile key that you're referring to), which can then be encrypted with any PIN, or any replacement for that PIN, and safely stored in memory after being encrypted yet again by the TPU via its burned-in random key. If not literally, then in effect. So ultimately, the situation is the same: extract the state of the TPU by atomic ablation, create a simulation of it, and run through all the (short) possible PINs.
I'm pretty sure I read about a shortcut to all this which was employed in a sub-million-dollar hack. They just found a wire they could ground that would prevent the chip from increasing the delay after each PIN was tried, and then brute forced it until it unlocked the phone. One would hope that modern TPUs aren't quite so daft. I don't think they are.