Hi Peter,

Thank you for reading the draft and for these comments. Ben Schwartz
has commented on the same draft on this list, and I have tried to
keep my answers to the two of you consistent.

> On 11/5/26 22:35, Shigeya Suzuki wrote:
>> - OPKEY, a new RR type with the TLSA RDATA format, restricted to
>> usage 3, selector 1, and matching type 1 or 2 [...] Until a type
>> value is allocated, the same RDATA is published as TLSA at the same
>> names.
>
> Why not stay with TLSA?
>
> In your underscore name specification, you can define semantics that
> expect usage 3, selector 1 etc.

We separated OPKEY from TLSA following the precedent of SMIMEA
(RFC 8162) and OPENPGPKEY (RFC 7929), which have types of their own,
so as to keep a record that means "this key is current for this name
in the Originator Profile Framework" apart from the TLS association
that TLSA means. That said, if TLSA under _op is acceptable to the
group, that would be fine with me as well; your point that the
semantics can be carried by the underscored name is well taken.

The SPF history you cite below points the same way. The draft's
transition was designed with that history in mind: it asks for the
type value at the outset, and confines the TLSA form to the period
before the allocation, so that the two forms are not kept in parallel
indefinitely.

The next revision will record the choice between a new type and TLSA
as an open question. I would be glad to hear from others, operators
of DNS hosting in particular, on which they would find easier to
support.

>> - An SVCB RRset at `_op.<originator>`, naming the issuers the
>> originator accepts. An AliasMode record with a TargetName of "."
>> states that the originator has no profile.
>
> That should be a PTR record.

Ben made a related point: SVCB identifies endpoints, and this record
names none. The next revision will move this record off SVCB, and PTR
is now among the candidates, with a dedicated type as Ben suggests,
TXT, and URI (RFC 7553). PTR has clear attractions here: it is an
existing type, its RDATA is a domain name and needs no parsing, and
an originator that accepts several CP issuers publishes several RRs.
What it would still need is a way to say that the originator has no
profile at all, which the AliasMode "." of SVCB gave us; I would be
interested in your view on how best to express that with PTR.

What weighs most in the choice is that this record is published by
the originator, typically through a commercial DNS hosting provider,
so a type that such providers let their customers publish matters
more here than it does for the issuers' records, which we assume are
published by a few flexible operators able to run DNSSEC.

>> - A TXT RRset at `<label>._revocation._op.<pa-issuer>`, whose
>> presence revokes an annotation on a profile.
>
> Your threat model apparently includes a DNS spoofing, else you would
> not need DNSSEC. (I think that's a reasonable threat model.)
>
> However, an attacker spoofing responses (at least when on path) may
> be able to suppress responses and instead replay old ones, which will
> validate unless their signatures have expired. This includes
> NSEC/NSEC3 records proving the absence of a revocation record.
>
> Your draft says on "Profile Annotations (PA)": "a verifier checks
> whether the PA issuer has revoked the PA ahead of the end of its
> Signature Validity Period".
>
> Under your threat model and with the previous paragraph in mind, an
> attacker may be able to hide the revocation and force the verifier's
> view to be stuck at some previous but still-validateable state. Only
> once signatures have expired, you can be sure the verifier has an
> updated view, but in that case you may as well have removed the
> annotation instead of publishing a revocation record.
>
> Hence, the value of revocation records is unclear.

Agreed. There are two bounds on when a revocation is seen. Without an
attacker, it is seen once cached denials expire, that is, within the
negative caching period of the zone (Section 10.4). With an attacker
on the path, it can be hidden for as long as the signatures on the
earlier answer remain valid; the RRSIG validity period is that bound
(Section 11.6). The same holds for the other records: a replayed
positive answer at an Originator Key Record, or at an issuer's key
record, hides the revocation of a profile or the withdrawal of a key
in the same way.

These bounds are meant to be read in the context the framework is
designed for. The annotations we have in mind attest, for instance,
that the originator is a member of an association, and the current
discussion assumes that such an annotation is reissued about once a
year. For them, a revocation that takes effect within a day or so in
the ordinary case is enough; against an attacker on the path, the
bound remains the RRSIG validity period, as above. The Core
Specification of the framework describes these windows in its
security considerations ("Revocation Latency and Replay Windows").

That is also where the record earns its place: an annotation, once
issued, travels with the content, and its issuer cannot remove it
from the copies already handed out, so without the record it would
stand until it expires. The next revision will state both bounds in
Section 8.3 itself, not only in the operational and security
sections.

> As for DNSSEC, I'm not concerned to the same degree as Ben. The
> draft doesn't say who a verifier typically would be, so there's no
> reason to assume it would not have DNSSEC validator available. Even
> if on an end-user device, the verifier tooling could include a
> DNSSEC validation library. (Section 11.1 actually says this, "web
> media use case [...] has a verifier carry a validator".)

That is also how I answered this part of Ben's message. Whether this
is practical in browsers is one of the observations the experiment
expects to make (Appendix B). What DNSSEC asks of deployment
otherwise is signing, and the draft asks it of the issuers' zones,
not of every originator's (the email use case being the exception,
as I mentioned to Ben).

>> wildcards are not allowed at these names.
>
> Why?
>
> Also, you can't guarantee that, as online signers often hide them
> (synthesized responses look like non-wildcard, which can have various
> operational advantages).

At these names, presence and absence are the signal. A wildcard at or
below _originator-key._op.<cp-issuer> would make an Originator Key
Record appear at every name below it, and the CP issuer could no
longer revoke one CP by removing its record; a wildcard below
_revocation._op.<pa-issuer> would mark every PA revoked (Sections 3.4
and 8.3).

You are right that this cannot be enforced from the verifier's side,
and the draft does not try: verifiers are not asked to tell a
synthesized answer from another. The rule is on the issuer that
publishes these records. A wildcard there is a misconfiguration, and
the responsibility for it is the issuer's; it does not only hurt the
issuer, since below _originator-key._op it makes revocation fail open
for every verifier. Section 3.4 gets two things wrong on this point:
it says that DNSSEC lets a validator recognize a synthesized answer,
which online signers make untrue in general, and that the harm of a
wildcard falls on the record publisher that placed it, which
understates the harm. The next revision will correct both.

A related accident is a wildcard at the apex of the issuer's zone,
which would reach these names if the _op node did not exist; keeping
that node in existence, as the issuer's own key record at
_issuer-key._op does, prevents it.

>> The draft is intended as Experimental. The experiment is whether a
>> new RR type can be put into service by the hosting providers and
>> verifiers involved,
>
> For context (but not necessarily with a pessimistic outlook),
> consider SPF, which, during experimentation, used specialized records
> of type 99 ("SPF") that mimicked TXT. This was later dropped in favor
> of TXT, see RFC 7208 Section 3.1.

Thank you; see my answer on TLSA above.

best regards,

shigeya

On Mon, Nov 9, 2026, at 02:18, Peter Thomassen wrote:
> Hi,
>
> (speaking for myself)
>
> On 11/5/26 22:35, Shigeya Suzuki wrote:
>> - OPKEY, a new RR type with the TLSA RDATA format, restricted to usage 3, 
>> selector 1, and matching type 1 or 2 (a digest of the SubjectPublicKeyInfo). 
>> It is published at `_issuer-key._op.<issuer>` for issuer keys, and at 
>> `<label>._originator-key._op.<cp-issuer>` for each originator profile. The 
>> presence of a Secure RRset means the profile is not revoked; a Secure 
>> NXDOMAIN or NODATA means it is revoked. Until a type value is allocated, the 
>> same RDATA is published as TLSA at the same names.
>
> Why not stay with TLSA?
>
> In your underscore name specification, you can define semantics that 
> expect usage 3, selector 1 etc.
>
>> - An SVCB RRset at `_op.<originator>`, naming the issuers the originator 
>> accepts. An AliasMode record with a TargetName of "." states that the 
>> originator has no profile.
>
> That should be a PTR record.
>
>> - A TXT RRset at `<label>._revocation._op.<pa-issuer>`, whose presence 
>> revokes an annotation on a profile.
>
> Your threat model apparently includes a DNS spoofing, else you would 
> not need DNSSEC. (I think that's a reasonable threat model.)
>
> However, an attacker spoofing responses (at least when on path) may be 
> able to suppress responses and instead replay old ones, which will 
> validate unless their signatures have expired. This includes NSEC/NSEC3 
> records proving the absence of a revocation record.
>
> Your draft says on "Profile Annotations (PA)": "a verifier checks 
> whether the PA issuer has revoked the PA ahead of the end of its 
> Signature Validity Period".
>
> Under your threat model and with the previous paragraph in mind, an 
> attacker may be able to hide the revocation and force the verifier's 
> view to be stuck at some previous but still-validateable state. Only 
> once signatures have expired, you can be sure the verifier has an 
> updated view, but in that case you may as well have removed the 
> annotation instead of publishing a revocation record.
>
> Hence, the value of revocation records is unclear.
>
> As for DNSSEC, I'm not concerned to the same degree as Ben. The draft 
> doesn't say who a verifier typically would be, so there's no reason to 
> assume it would not have DNSSEC validator available. Even if on an 
> end-user device, the verifier tooling could include a DNSSEC validation 
> library. (Section 11.1 actually says this, "web media use case [...] 
> has a verifier carry a validator".)
>
>> wildcards are not allowed at these names.
>
> Why?
>
> Also, you can't guarantee that, as online signers often hide them 
> (synthesized responses look like non-wildcard, which can have various 
> operational advantages).
>
>> The draft is intended as Experimental. The experiment is whether a new RR 
>> type can be put into service by the hosting providers and verifiers involved,
>
> For context (but not necessarily with a pessimistic outlook), consider 
> SPF, which, during experimentation, used specialized records of type 99 
> ("SPF") that mimicked TXT. This was later dropped in favor of TXT, see 
> RFC 7208 Section 3.1.
>
> Best,
> Peter
_______________________________________________
DNSOP mailing list -- [email protected]
To unsubscribe send an email to [email protected]

Reply via email to