Thanks for the clarification.  It sounds like the intent here is that:

1. The SVCB signal is not included in referral responses.
2. The SVCB signal is acquired during NS revalidation.
3. This avoids the additional work for resolvers to probe port 853.

I can understand the logic of this, but most resolvers do not perform NS 
revalidation, so stacking this capability on top of NS revalidation would limit 
the number of resolvers that are able to deploy this behavior.  Also, RFC 9539 
probing can generally run in parallel to NS revalidation, so it should be 
faster than waiting to start TLS until after receiving the SVCB hint.

--Ben


________________________________
From: Johan Stenstam
Sent: Wednesday, June 25, 2025 5:27 AM
To: Joe Abley; Ben Schwartz
Cc: Working Group DNSOP; [email protected]; Erik Bergström; Leon Fernandez
Subject: Re: [DNSOP] Proposal for opportunistic transport signaling from 
authoritative servers

Hi,

Many thanks for the comments. There are a bunch of things to sort out. Let’s 
start with the issue of whether this is worthwhile or we should just wait for 
DELEG.

Using an opportunistic scheme like this is not able to defend against an active 
attacker. The complete solution is clearly DELEG. Full stop.

My concern with that is that DELEG will not be able to add privacy to a large 
fraction of the public zones for, well, a rather long time. The reason is that 
even if the specification was complete tomorrow, and implementations were done 
next week we would still have to wait for all the registries and registrars to 
add DELEG support. And after that every single zone will need to be upgraded 
which will also take years.

I’m not objecting to this, there is just no way around the fact that a major 
change to the DNS like DELEG will take a long time to reach wide scale 
deployment. Just like DNSSEC. Or IPv6. Or...

The point with our proposal is that because it is so simple, and because it 
doesn’t require any changes to the registry/registrar/registrant ecosystem 
(only providers of authoritative or recursive services need to do anything) 
this could, if we want, reach wide scale deployment in a fraction of the time.

I argue that we want to add more privacy to authoritative DNS now and the 
reason we have not done that yet is that we have not (AFAIK) had a sufficiently 
simple mechanism to provide transport signaling from auth to resolver within 
the current protocol. And blind probing a la RFC9539 was not not sufficiently 
attractive (as the odds of it working were low).

This proposal is just an opportunistic hint that says “Hey! I support DoQ, 
please try it” to allow the resolver to do something like RFC9539, but with 
open rather than closed eyes and therefore a vastly larger chance of success.

On 24 Jun 2025, at 21:46, Joe Abley <[email protected]> wrote:
On 24 Jun 2025, at 21:38, Ben Schwartz 
<[email protected]<mailto:[email protected]>> 
wrote:

1. An attacker could strip the SVCB record and its RRSIG, resulting in an 
ordinary delegation response that would be accepted and used without encryption.

True. A client that didn't find a record could query for it which would give it 
the ability to validate any non-existence, but that sounds like a lot of 
pointless queries in the situation where the majority of zones don't have this 
kind of thing deployed (when most referrals don't include a courtesy SVCB).

All of this is true and I am not suggesting actively querying for the SVCB. If 
the resolver gets the transport signal for free it may use it, that’s all.

2. This is a child-side record, and so presumably also a child-side RRSIG, but 
the validator may not yet have the child's DNSKEYs.  Fetching those keys (in 
order to validate the transport configuration) would create additional delay 
before the query could be sent.

That additional delay can be amortised over the lifetime in the cache of those 
records needed to provide a chain of trust though. A validator is going to have 
to do that work at some point.

I agree to the amortization point. However, I want to point out that that we’re 
not talking about the child’s DNSKEYs, we're talking about the DNS provider’s 
DNSKEYs. I.e. for the child zone "example.parent.”, with the DNS provider 
Cloudflare the response would be something like:

…
Answer:
www.example.parent.  IN A 1.2.3.4
Authority:
example.parent.      IN NS ns.cloudflare.com.
Additional:
ns.cloudflare.com.   IN A 5.6.7.8
ns.cloudflare.com.   IN AAAA 2001::bad:cafe:53
ns.cloudflare.com.   IN SVCB 1 . alpn=doq,dot
ns.cloudflare.com.   IN RRSIG SVCB …

And the likelihood of resolvers already having the DNSKEYs of all the major DNS 
providers is rather high.

3. The parent-side logic to serve this RRSIG, and the coordination channels to 
get it there, seem complicated.

I guess. It seems approximately as complicated as CNAME flattening which is 
pretty common.

This part I don’t understand. There is no “parent side logic”. There is no 
change whatsoever to the delegation information (otherwise this would be a 
non-starter).

The proposal is to add the transport signal when sending an authoritative 
response, i.e. an Answer, from an authoritive nameserver (where the auth 
nameserver will make a statement about its own transport capabilities).

There is no suggestion about changing anything in a referral from the parent.

Assuming it is signed by the ZSK, this also breaks a usual rule of not having 
parent-side content depend on the child ZSK.

The referral can be processed without it (by using other transports) so I'm not 
sure it's a dependency.

The downgrade attack seems worth thinking about. I'm not sure how much we need 
to care about the other things.

I think there may be a misunderstanding hidden here. But as I don’t understand 
what you’re talking about here I may be wrong…

Regards,
Johan

_______________________________________________
DNSOP mailing list -- [email protected]
To unsubscribe send an email to [email protected]

Reply via email to