Re: [blink-dev] Re: Proposal: Marking HTTP As Non-Secure

On Fri, Dec 26, 2014 at 11:36 PM, Ryan Sleevi <rsleevi@chromium.org> wrote:

(snip)

> >>
> >> That is, the standard could have provided policy and the site could
> >> have sent a policy that governed the pinset. The site could have
> >> allowed an override or denied an override. But it was decided all
> >> users must be subjected to the interception, so the policy elements
> >> were not provided.
> >>
> >> So how's that for strategy: the user does not get a choice in their
> >> secure connection and the site does not get a choice in its secure
> >> connection. Rather, some unrelated externality makes the choice.
> >
> >
> > I share your concerns, and more: Public Key Pinning (and Strict
> Transport Security, for that matter) is awfully specific to HTTP, even
> though it has nothing to do with HTTP at all. The proper place to put this
> information is in DNS records. This has worked reasonably well for Email;
> or at least I don't see anyone proposing we query SMTP servers for relevant
> security metadata.
>
> I assume that you aren't familiar with STARTTLS? The encryption and
> security story for email is disastrously worse than anything for HTTP.
>
> > So why do this for HTTP? (If one wants to answer that properly stopping
> forged headers is less important than stopping plaintext cat pictures,
> color me surprised; if it's because it's a better alternative than
> unencrypted DNS records for those lacking DNSSEC, that's a better answer,
> but it's still relegates HPKP and HSTS to the realm of "stopgap measure".)
> >
> > Austin.
>
> Indeed. From the point of view of client applications, DNSSEC is a
> complete and utter failure, and will remain so for the next decade, given
> hardware cycles. If a decade sounds like a stopgap measure, when it was
> longer than the relevant lifetime of sites like Myspace, so be it, but that
> seems like a pretty good run to me.
>

I would challenge this notion: The deployment cycle doesn't seem to be that
much longer than TLS (underneath HTTP). SSL was first available in 1995,
standardized 1996, almost twenty years ago. TLS 1.0 came out 1999 and we
STILL have user-agents that will happily downgrade to SSL 3.0. (I don't
claim to be a TLS/DNSSEC historian, feel free to correct me.)

While efforts to secure DNS began about the same time, the modern DNSSEC
was first operational around 2004, and the first Key Signing Ceremony was
only in 2010! I find it fully usable, for those applications that support
it. At the application client level, I haven't found any reasons to not
support it.

I would venture to guess that Web browsers and other user-agents could make
a better impact by deploying DNSSEC support, over forcing TLS usage. The
former adds more options for security; the latter imposes costs on users of
TLS, whether or not the existing system suits their needs.

I don't see anything wrong with STARTTLS; in fact I find it preferable to
having seperate URI schemes for what is otherwise the same resource. (I
don't actually run any email servers, but I use STARTTLS with LDAP, XMPP, a
proprietary protocol, and the similar RFC 2817 for HTTP. That is, in fact,
a thing.) If one requires security, they can always start requests with
"STARTTLS"; if a server demands security (for its definition of security),
then it can just kill plaintext queries. Nor do I find deployment of TLS on
email systems to be particularly behind TLS on HTTP (do you have data on
this?). In particular, I do like Alex Russell's assertion that TLS is about
"quality of service" (if I understand correctly; my response point being
this is not exclusively what security entails), and this is something that
exists below the application layer.

Austin.


> >
> >>
> >>
> >> Jeff
> >>
> >> On Fri, Dec 26, 2014 at 8:33 PM, Austin William Wright <aaa@bzfx.net>
> wrote:
> >> >
> >> > On Wed, Dec 24, 2014 at 3:43 PM, Alex Russell <slightlyoff@google.com
> >
> >> > wrote:
> >> >> ...
> >> >>
> >> >> No, the request for a resource means something. Security is about the
> >> >> quality of service in delivering said thing.
> >> >
> >> > Specifically, I refer to privacy and authentication (which seems to
> be what
> >> > "security" means here). There's many components to security, and
> they're not
> >> > all going to be desired at the same level for every request,
> sometimes even
> >> > at odds with each other. Deniability is a particularly expensive one
> to
> >> > implement, often at significant monetary and performance cost to the
> user.
> >> > Deniability is simply not implemented by any website (or more
> accurately,
> >> > any authority) indexed by the major search engines.
> >> >
> >> > It's difficult to claim that deniability, though, is about "quality of
> >> > service", but it is nonetheless considered something _very_ important
> by
> >> > Tor, and by users of it.
> >> >
> >> >>
> >> >> > HTTP could mean, simply, "I don't care about security".
> >> >>
> >> >> Then user agents should be free to act on behalf of users to write
> off
> >> >> correspondents who do not value the user enough to protect the
> message in
> >> >> transit.
> >> >>
> >> >> You're making claims about what UA's must do from the perspective of
> >> >> servers; this is a clear agency problem misunderstanding.
> >> >>
> >> >> We're the user's agent, not the publisher's.
> >> >
> >> > When a user navigates to an https:// resource, or hits the checkbox
> for
> >> > STARTTLS, or selects "Require encryption" from a drop-down, they, the
> user,
> >> > are *demanding* a secure connection (for someone's definition of
> secure;
> >> > traditionally this means privacy and authentication but need not
> include
> >> > deniability).
> >> >
> >> > If we want to talk about the perspective of servers, the server has
> options
> >> > to demand a secure connection too; it simply has to deny plaintext
> requests
> >> > (it could redirect, but this risks the user-agent sending sensitive
> >> > information including request-uri that'll just be ignored), and TLS
> can
> >> > specifically ask for or require a client TLS certificate for
> authentication.
> >> >
> >> > The user-agent works for the user, not the other way around. Letting a
> >> > user-agent intervene in the case of a plaintext request is imposing a
> value
> >> > on the user: It prioritizes privacy over "DO NOT BREAK USERSPACE!"
> which is
> >> > not *necessarily* true.
> >> >
> >> > If TLS were free or very cheap to implement, however, this would be a
> >> > non-issue. Hence my treatment of ways to do this.
> >> >
> >> > Even if the http-as-non-secure proposal fully worked; it still
> increases the
> >> > cost of publishing content and barriers to entry. I recall Tim
> Berners-Lee
> >> > recollecting that requiring users mint a long-lived URI was
> concerning cost;
> >> > but there was no other way to design a decentralized Web. (In the
> end, the
> >> > biggest barrier to entry has been acquisition of a domain name and
> hosting.)
> >> >
> >> > Of course, if efficiency and security are at odds, then we prefer
> security.
> >> > However, I see making TLS implementation cheaper as a viable
> alternative to
> >> > this proposal, and so is far more appealing.
> >> >>
> >> >> > A TLS failure means "Security is _required_ by the other party, but
> >> >> > couldn't be set up and verified. Abort! Abort!"
> >> >> >
> >> >> > Purporting that plaintext and HTTPS failure is the same would be
> >> >> > conditioning users to be less worried about HTTPS failures, where
> there's a
> >> >> > known _requirement_ for confidentiality.
> >> >>
> >> >> The UA could develop different (but also scary) UI for these
> conditions.
> >> >
> >> > I am specifically referring to the part of the proposal that, in the
> long
> >> > term, specifies no UI difference between plaintext and failed-TLS.
> >> >
> >> > Would you care to make an alternate proposal with this change?
> >> >>
> >> >> Regardless, we have done a poor job communicating the effective
> system
> >> >> model (that content is not tamper-evidentwhen served over HTTP).
> Steps to
> >> >> make this clearer aren't the same thing as throwing all distinction
> >> >> overboard.
> >> >>
> >> >> > There may be some value in saying "as with many network requests,
> >> >> > request/submission is being sent in the clear. Anyone on your
> network will
> >> >> > be able to read this!". But big scary warnings are most certainly
> bad. No
> >> >> > matter what, before I claim anything for sure, I would like to see
> a
> >> >> > double-blind study, and not change for the sake of change.
> >> >>
> >> >> This sort of argument might have worked in '12. Apologies, but you're
> >> >> dreadfully late.
> >> >
> >> > I'm afraid snark wasn't my first language, so you'll have to refresh
> my
> >> > memory: what are you referring to?
> >> >
> >> > Earlier in my message, I touched on the danger of users dismissing
> warnings
> >> > for their e.g. bank, because users become dismissive of "insecure
> >> > connection" warnings found elsewhere.
> >> >
> >> > Now, we might have HSTS (for my bank example), but at that point
> we're no
> >> > longer talking about two tiers of security, but three
> (insecure-bypassable,
> >> > insecure-nonbypassable, secure), which is precisely what the proposal
> is
> >> > advocating moving _away_ from (or, the proposal is not very clear on
> this
> >> > point, and a clarification in its text would be necessary).
> >> >> ...
> >
> >
>