Telechat Review of draft-ietf-privacypass-auth-scheme-12
review-ietf-privacypass-auth-scheme-12-httpdir-telechat-thomson-2023-08-29-00

This document describes a new authentication scheme for HTTP that enables
authorization via Privacy Pass.

I've been tracking this work at a high level, but I've never reviewed this
document in any detail until now.  After taking a closer look, I've found quite
a few problems.

https://github.com/ietf-wg-privacypass/base-drafts/issues lists 23 new issues. 
https://github.com/ietf-wg-privacypass/base-drafts/pull/459 suggests some
editorial fixes on top of those.

A number of those issues are significant enough to suggest that the document is
not ready.  I expect that most will be easily handled, but there are a couple
of trickier ones.

https://github.com/ietf-wg-privacypass/base-drafts/issues/448 is serious enough
to draw special attention to.  In short, Section 3 of the document is very
problematic as it makes a lot of assumptions about a particular deployment
environment.  Some of those assumptions are -- I think -- bad.  It looks like
the intent of this section is to describe how this mechanism might be deployed
safely to the Web.  This raises a number of concerns:

1. This is an IETF document.  The W3C is probably in a better position to come
to conclusions about what is (or isn't) appropriate for deployment to the Web.

2. The bounds on user agent behaviour are not specified in sufficient detail. 
There is definitely a case to be made for this to be deployed to the web as
envisaged, with different implementations making their own choices.  But if the
intent is to describe the nature of the risks involved in deployment to the
Web, then there is not enough detail to guide the successful implementation and
deployment of this feature.

3. There are a number of implicit assumptions throughout that are not
adequately explained.  For instance, there is discussion of use of this
mechanism across origins or sites, despite that violating established Web
norms.  There is discussion of that cross-site transfer occurring without user
involvement, which is a oft-used safeguard against such privacy leaks.  But the
necessary preconditions for that transfer are not articulated.  There are
potentially scenarios where this sort of transfer could be safe, but there are
great many where it is absolutely not.  The document seems to be assuming that
the token carries a very particular signal along with it, namely that the
client acts on behalf of an entity that the attester (and transitively, the
issuer) believe not to be abusive.

I've suggested in the issue that this section needs considerable revision. 
There are general requirements on the use of the protocol that are currently
buried in amoungst Web-specific requirements.  Those will need to be teased
out.  It's possible that the accompanying architecture document could cover
this material, but I'm not seeing it there.

Then there are the Web-specific requirements that really belong in a
Web-specific document, which is probably something that the W3C is in a better
position to produce than the IETF.  Here, the precise set of safeguards will
probably be some mixture of client-specific policy and widely-agreed policy, so
getting that mix right will need careful consideration.

Despite all this, I'm generally supportive of this protocol.  It is a design
that should work well in a great many contexts, but getting this right for the
Web requires more than this document provides.