Neither HPKP nor HSTS deal with the issue of a-man-in-the-middle doing a downgrade attack against clients who are visiting a site for the very first time (which can be encouraged through phishing). There are only three ways to deal with this that I can think of:
1) Deprecate insecure HTTP. Literally get all the browsers to agree on a deadline X years out and then drop support for it entirely. HSTS would become irrelevant. I hold little hope of this ever happening, even in a HTTP/2 world. See IPv6
2) Keep increasing the size of the trusted, preloaded list of HTTPS enabled sites shipping with browsers: Neither the implementation nor the process to manage the list scales. It's just crappy all-round.
3) Standardise some SRV records for web protocols in the DNS so browsers can efficiently determine what is supported: Requires DNSSEC to prevent downgrade attacks.
Of these, #3 seems to me to be most practical. With regard to DNSSEC issues I would give these counterpoints:
* DNSSEC could and should be augmented to support ECC, obliterating the concerns with regard to large records and weak RSA key lengths.
* Browsers (or your OS's dnscache process) can cache DNS responses from most of the chain. If the ZSK hasn't been rolled over for a particular domain then I don't see the need to reverify it again on each request. It's not like revocation really works in HTTPS today anyway, so you probably aren't going to bother to check upstream for NSEC records on each request. Or just recheck upstream once a day. I don't really know, but I'm sure these latency issues can be minimized. Googles obsession with latency is how we ended up with the barf that is HTTP/2 rather than something that actually improves websec in the first place.
Another thing I don't often see people mention about HSKP and HSTS, regardless of whether they can be used as "supercookies", is that they create a more-or-less permanent shadow browser history on your machine. This alone is borderline show-stopper afaiac
Whether or not DNSSEC is ever seriously deployed, downgrade attacks against HTTPS will remain viable. It's easy to see one reason why this is true (DNSSEC only protects the DNS lookup and not the HTTP traffic itself), but there are other downgrade attacks as well. Downgrades are hard to protect against.
The objective is to leverage DNSSEC to prevent downgrade attacks. Whether or not you are connecting to a host over HTTP or HTTPS you need to query DNS, which makes it an ideal place to introduce a message to clients to tell them what they should be doing. In an ideal world, a DNSSEC enabled recursive resolver, querying a domain name under a DNSSEC-enabled TLD, should not be fooled to downgrade.
Again: this mitigates (but probably does not decisively solve) one avenue for downgrade attacks. But downgrade attacks against HTTPS would remain possible --- trivial, in fact, without HSTS, which leaves you with the first-contact problem with or without DNSSEC.
So again: what's the point? Compared to HSTS headers, DNSSEC is incredibly expensive.
The point is you can have a record in DNS that says "never use HTTP" that protects clients who haven't yet seen HSTS, and you can use a TLSA-like SRV record to protect clients who haven't yet seen HPKP.
Another thing you're missing is that the CA system almost boils down to the integrity of DNS already, since you can get a CA to issue a basic certificate for a domain simply through weak ownership verification (i.e. if someone controls or MITMs your MX records/responses you're fucked).
Once again: you can foil domain validation with or without DNSSEC by downgrading SMTP.
Why bother with DNSSEC?
You also didn't address my ECC point upthread. Today, APNIC advises DNS administrators not to use the (crappy) P-256 ECDSA DNSSEC supports, because it breaks ~1/3rd of all resolvers. That's for the ECC variant DNSSEC actually "supports".
How exactly would you propose a rollout of Ed25519 (or equivalent) crypto in DNSSEC? Or, where is the flaw in my argument?
4) Ask sites to publish via an out-of-band channel their certificate fingerprints in an easy-to-verify manner.
I think the centralised CA model really needs to be replaced, although it will be hard to displace as long as those involved have a financial interest in continuing it.
The problem you're describing has been donned Zooko's triangle. It's unclear whether decentralised solutions to authenticated human-meaningful names can ever scale to something as large as the DNS.
While adoption of DNSCurve would of course be wonderful for privacy online, it cannot be used right now to provide DNS record authentication. The root and TLD nameservers would need to enable support for it, otherwise you'll always be able to MITM the upstream NS record response and redirect all further queries to your own server. Given the politics involved in DNSSEC, and the current anti-crypto climate, I'd say the chances of a fully-chained DNScurve deployment ever happening are about absolute zero.
No, DNSCurve doesn't authenticate DNS records, it only secures integrity and privacy of your requests (and the responses). DNSCurve and DNSSec don't solve the same problem, they're actually complementary.
I agree with his comments on HPKP. I looked in to adding HPKP headers to a couple of my sites, and figured out how to do it, but I'm nervous about enabling it. It seems far too easy to make a mistake and lock people out of being able to visit your site. The trouble is, if you make a mistake, they're not locked out until you get around to fixing it. They're locked out until the expiry date which you set in the HPKP headers, which could be months away.
Tools to permit you to wipe out your HPKP mistake would be usable by attackers, though. That's hard to solve.
I haven't looked into HPKP much, but it appears it has the ability to be configured in a "report only" mode if name your header "Public-Key-Pins-Report-Only". This will allow you to test your implementation to make sure it's correct before renaming the header to start enforcing.
The idea of a non-enforcing mode for testing purposes also exists in DKIM (though DKIM didn't really catch on much due to servers with proper DKIM configurations getting compromised and sending spam...)
I would say that HSTS is far more dangerous as I don't see a similar "testing" mode available.
edit: It would be a good idea to make sure your max-age setting doesn't exceed the expiration date of your intermediate certificate, so don't go crazy with it. A tool with built-in safeguards to handle most of this configuration would be welcome.
"Tools to permit you to wipe out your HPKP mistake would be usable by attackers, though."
I meant tools to prevent you from making the mistakes in the first place, not for fixing it after the mistake.
The non-enforcing mode doesn't help. When I talk about mistakes, I mean things like losing access to your cert or timing mistakes in rolling the cert. If I roll it out today, it will be fine today. It will probably fail at some point in the future though, and fail hard, unless I follow very strict processes.
HSTS is considerably less dangerous. Once it's set up, it will continue to work without you having to do anything extra. HPKP on the other hand requires you to actively consider when you generate new certs, and when you add extra HPKP headers to your responses, and when you roll the certs etc.
I can change my CA or generate a new key today on a whim, because I don't do HPKP. If I did do HPKP, this would require considerable thought and preparation in advance.