No, it's security by intrusion detection, generally. HSMs are designed to be boxes that it's very hard to get a secret out of with physical access. TPMs generally aren't the most paranoid version, but it gets more expensive and less practical as you go further (e.g. a large box which has a battery backup, keeps the secrets in RAM, and will wipe them as soon at it detects any funny business. These are DIYable, but the list of tricks by attackers is long and it's hard to cover all of them at once). A TPM is generally somewhere in between that and a regular micro with no particular effort to prevent readout of internal storage, in that they are small, can persist secrets without power, but are still difficult to attack physically (~maybe at the level of advanced criminal organisations, ~probably at state level if they're willing to spend some money on it, even absent a backdoor).

Unlike with cryptography, there is no rigorous notion of physical security. Doors, locks and even security systems can all be overcome with sufficient effort, skill and resources. They work because physical attacks require proximity and are very hard to keep anonymous. I seriously doubt that any TPM implementation would last a week against government funded researchers with state of the art technology, but that doesn't mean the TPM is useless.

No, it's the same. Cryptography is like a lock that you can overcome with mathematical force. It's just in different domain than physical objects.

If you know how the lock is built, you can rule out existence of master key for instance. You don't know if your TPM chip has API where three letter agency can just download the keys from it. You are in the dark.

Same with cryptography, you can choose the method, just like you can choose type of lock. There are locks that have not yet been picked, but you can use a hammer, similar with cryptography - you can use quantum computer etc.

Which locks haven't been picked? Abloy Protec 2 got picked, Bowley got picked, StealthKey got picked… I'm not aware of any designs for an unpickable/unbypassable lock. Whereas several AEADs have not been broken.

Some of these features makes it harder to (physically) probe the internal parts of the chip and read out secret values:

https://en.wikipedia.org/wiki/Secure_cryptoprocessor#Feature...

These things make it harder to break into the internals of the chip regardless of they being kept secret, so I wouldn't call it security by obscurity. I'm not even sure you can apply that principle to physical security.

Printing out your 2FA recovery keys and storing them in a safe also is.

Is a yubikey then employing security by obscurity?

100%

They’re built from essentially the same secure MCUs as traditional TPMs and both the hardware and the proprietary crypto libraries used on them have been exploited many times over.

But would you not agree that using a yubikey can improve security? If you chose to label TPMs as security by obscurity, so be it, but it doesn’t make them less useful conceptually. Shitty implementations and complexities of UEFI do that.