Since we already are at the point where the security of the system hinges on the registrar, why not at least limit the damage potential by only accepting certificates issued by the registrar in question? Instead of letting any CA under the sun ALSO issue these?
PS: If certificate issuance was a required and expected task for all registrars, it could even lead to $0 certificates, which would surely boost HTTPS usage. It would even provide an incentive to pick a thorough registrar if you are concerned about security and validation, since no other entity could issue illegitimate parallel certs behind your back. Brings some meaning back to being a high quality authority (even charging more for more trust) instead of the race-to-the-bottom we have today. Sloppy registrar-CAs would lose business because their trustworthyness actually means something for once. (You wouldn't register your serious domain with a registrar known to have poor validation checks in place for issuing certificates, for example)
You seem to have gotten things backwards. If the DNS is hijacked, SSL is the only thing protecting a visitor from being lured into the hijacked site.
Connecting these two things into one makes SSL effectively worthless.
They way it is now, hijacking DNS (between the domain NS servers and any CA) would allow an adversary to request and obtain an illegitimate certificate by way of domain validation. If the registrar is the only valid issuer of certificates, this loophole is closed since there is nothing for the CA to verify - the registrar already knows who the customer is and can offer certificate signing without an insecure DNS based domain validation.
For end-users with a hijacked DNS, the registrar-issued SSL certificate would still protect the transmission, because the DNS-spoofing adversary would not be able to present a valid certificate signed by the registrar. In fact, in today's CA environment, if the DNS hijacker is playing ball with a rogue CA, they could also spoof the SSL. With my suggestion, they couldn't, unless the rogue CA is actually the registrar, because the browser would see that the SSL certificate was not signed by the appropriate registrar.
You could create a DNS-style CA setup like this:
* Browsers would ship with ONE root CA, the public key of the "." root zone operator
* Each TLD ("com", "org" etc) has a CA signed by the root zone operator, valid only for signing TLD registrar CAs. The root zone CA could publish a signed list of TLD certificates daily/weekly/monthly (they shouldn't change too often and the number of entries would be relatively low). Anyone could compare notes and see if these change. There aren't any hidden intermediary CAs. Browsers could sync this list daily/weekly/monthly, the deltas should be minimal. ISPs could even provide mirror services, because the whole list is signed by the root zone operator anyways.
* Each registrar under a TLD has a CA certificate signed by the TLD operator, valid for only signing secondary domains under the given TLD. (i.e. customer domains). Just like the RootZone->TLD signed list of CAs, the TLD CA operator could publish a signed list of registrar certificates daily/weekly/monthly (again, they shouldn't change too often and the number of entries would be relatively low). Again, anyone can compare notes and see if things change. No hidden intermediary registrars. Browsers could yet again sync this daily/weekly/monthly and ISPs can mirror the list since it's signed.
* Registrants of customer domain ("example.com", "example.org" etc) can request a domain-specific CA at their registrar, and only at their registrar. This domain-specific CA is valid only for signing leaf domains, i.e. "www.example.com", "mail.example.com". With the domain-specific CA in hand, the customer can create their own leaf domain certificates "at home". The registrar should offer non-EV domain-CA certificates for free (there is no work to be done to validate domain ownership as the customer already has an account there), or charge a fee for EV certificates.
* There needs to be a way to determine who is the valid registrar for a given domain. This could be for example a TLS-based machine-readable WHOIS service address operated by each TLD. The TLS-based WHOIS service would negotiate TLS with the same TLD-CA-certificate as specified in the root list. Thus, a browser can connect to the TLD's WHOIS service and validate that a given domain is under a given registrar, and that the leaf SSL certificate is signed by the domain CA certificate, which is signed by the registrar CA certificate, which is in the signed-by-the-TLD list of registrars, which is in the signed-by-the-root-zone list of TLDs. This stuff could also be cached at the ISP level because everything is chain signed.
Maybe this is how DNSSEC works (I haven't looked into the details), but I think this would be a pretty neat way to limit the damage done by rogue CAs.
Not exactly, but it's very similar. DNSSEC has no concept of registrars, and each zone is entirely signed by the same entity.
You'd trust that single entity to inform the registrar anyway, so descentralizing it gains you nothing.
I would have to think about it but I'm sure there is SOME way to verify someone (either in person or digitally). This process would only have to happen once to verify the CA cert for their domain. After that - if their CA cert is compromised it's not a big deal because it only affects one domain.
Oh - and any hoster that allows it's employees to reset virtual machine passwords or make account modifications via email/support ticket should NOT be allowed to participate in this scheme.
Also, DNSSEC still has some severe issues in practice. I'd be glad if we could get DANE widely deployed for mail servers.
Edit: Oh, and if you're not using a validating resolver yourself you're also trusting that you're ISP is using one and not manipulating the responses.
I really don't understand why people keep repeating that complaint. Of course, if you don't check the keys you don't get any security.
How is that a problem of the algorithm? And how is that a problem on practice? If you want some real amount of security you check the keys, being them SSL certificates, DNSSEC signatures, or whatever else encryption system people put on place.
That's why DANE for mailserver is such an attractive target. They're usually run by people who know what they're doing and it helps bring a lot of infrastructure into place.
The fact that current software is hard of configure is just a symptom that it's badly designed. The only inherently hard thing in DNSSEC is distributing your domain data (not really harder than setting our server for TLS), and normal people do not do that.
I think the trust model is broken for the same reason it can be said three can keep a secret, if two of them are dead.
How can humans create a software or hardware system that absolves us of the issue of trust, when people are, and have always, been up to no good.
The courts are full of people who claim other people have acted outside of good judgement.
I don't think it's possible to have trust without its opposite.
I'd like to be proven wrong.