minor tweaks to the cert-chain section
I noticed the "delegate" field went away.. good idea. So I think signers will fall into two cases:
(what should the iss of the second (delegated-to) cert be? It feels like the only place where iss is relevant is in the very first cert, as it tells the verifier where to find the first key).
There's an asymmetry that feels slightly odd (but probably safer): both cases are really delegating authority to a second key, but allowChaining is required if the first key has principal.domain instead of principal.email, or if the second key wants to have a principal field instead of aud. The number of words needed to describe that makes the flag seem a bit artificial. Maybe the rule is "if allowChaining=false, the next certificate must be the last one (i.e. the assertion)".
What might seem cleaner? Here's a straw-man.. if we didn't use allowChaining, but instead relied on principal and exp to control delegated power, would it still be safe? It would mean the RP couldn't prohibit delegation: users could build arbitrarily long certchains, but they'd still be limited to their one principal.email=. And there's a relevant maxim from the obj-cap world (from Ka-Ping Yee) which says "don't prohibit what you can't prevent". Since users can sign as many assertions as they want (until exp runs out), prohibiting them from delegating to other keys isn't really any safer, but would annoy a user who wants to e.g. delegate to a cluster of per-browser short-expiry keys and keep their main key mostly offline (exactly analogous to the load-balancers scenario for the RP).
The extra flag is appropriately conservative and opt-in: nothing new happens unless it's present. But I'm looking for a way to define what exactly it enables that doesn't make the normal "enable aud=" case seem out-of-place.
minor wordsmithing on cert-chains
Note to self: This is still relevant and being held onto for future use, but not for immediate merging.
I thought we removed chaining when we discovered a security bug with them, no?
@fmarier We did, but because of a bug in our own implementation :) The idea itself is good, though probably not for a 1.0 version of the spec.
Francois and I were just talking about this one. All of the use cases we were able to come up with last year are now handled, without chaining, by various features we have added (or are planning to add) to Persona: the fallback's query arg for "which domain you actually asking for" used by bigtent, and putting multiple keys into the .well-known support document. Those features aren't too ugly.
So I'd suggest deferring chaining indefinitely, at least until someone comes up with a new use case that isn't easily handled by hacks of those magnitudes.