On Sun, Oct 18, 2026 at 1:46 AM John Levine <[email protected]> wrote:
> It appears that Wei Chuang <[email protected]> said: > >... I agree that we shouldn't underestimate the difficulties of > >introducing a new algorithm, as we saw with the friction of getting DKIM > >operators to adopt Ed25519. I've recently heard one reason was library > >compatibility issues. > > I don't think that's it. Everyone figured out how to handle RSA > signatures, > and the spec we wrote for Ed25519 was clear enough that nobody I know had > significant trouble implementing it. > My point is simply that we must respect the effort involved in adding a new algorithm. My other point is that the draft includes language requiring the specified algorithms to be mandatory, which raises a concern that this could become an adoption barrier if we don't consider it carefully. > > As I said: > > >> Key rotation in DKIM1 doesn't work in practice, mostly because you have > >> to have separate selectors for different algorithms which means double > >> signing, which risks losing mail in badly designed receivers that > >> penalize signatures they can't verify. I think we fixed this in DKIM2 > ... > > ... by allowing multiple signatures in each header. They have to be ORed > or algorithm rotation isn't going to work. I don't think an AND option > makes sense unless you can explain to me what is the correct thing to do > with a signature with two signatures, one of which you can verify and > passes, > and the other you don't understand at all and can't even try to verify. > For the multiple algorithm support I think there are three approaches to computing the overall verification results, along with rationale. We can evaluate the deployment pattern to see if that helps narrow down to a preferred approach. For each DKIM2 signature with multiple signatures, find the individual verification results, then use the given reduction approach to determine the overall verification result for the DKIM2 signature header field. * ORing of multiple signatures - to support buggy implementations with greater security risk when there is an algorithm vulnerability. * ANDing of multiple signatures- protection against algorithm vulnerability but has interop risk * AND/OR under the control of some sender specified flag- supports ORing during bring up and ANDing when security is needed though w/greater complexity. I would also argue the current -06 spec draft is closest to ANDing (despite the claims that the SHOULD acts as an OR) because the language says <https://datatracker.ietf.org/doc/html/draft-ietf-dkim-dkim2-spec-06#name-perform-the-signature-verif> : "If there is more than one signature provided then they MUST all be checked if the Verifier is able to do so. *If any signature fails then an error SHOULD be reported*. If all signatures that can be checked fail then FAIL MUST be reported." While the ambiguity is intentional, I believe the specification would benefit from greater clarity by picking one of these approaches and being very explicit about it. During algorithm bring up, the baseline algorithm is RSA and either ed25519 or mldsa44 and new implementations may have bugs. Since senders have a very strong incentive for deliverability, an AND requirement combined with a buggy implementation means authentication failures, resulting in negative deliverability. Senders under an OR requirement can deploy the old and new algorithms and use to-be specified monitoring to see that their new implementation is buggy. With the flag approach, the sender uses the OR'ed signature setting during bring-up and can switch to AND'ed afterwards for protection. I believe that when a Cryptographically Relevant Quantum Computer (CRQC) is imminent, we'll get warning. With the OR approach, the sender can switch to the PQC algorithm mldsa44. The AND approach is fully protected, so it doesn't have to change. The flag approach allows switching to the AND'ed signature setting if they have not done so already. If the vulnerability disclosure is a surprise, the vulnerability window is from the likely point of exploitation until the sender can change to an unaffected algorithm; however, as noted before, threat actors have easier ways to spoof email. With the sender specified flag, the default reduction approach is OR. However the sender can specify the flag "f=" with a value of "and", that causes the multiple signatures to have a reduction of AND. We would specify a new "f=" tag in the Spec section 8. <https://datatracker.ietf.org/doc/html/draft-ietf-dkim-dkim2-spec-06#name-f-flags> 10 <https://datatracker.ietf.org/doc/html/draft-ietf-dkim-dkim2-spec-06#name-f-flags> as "and". I don't think this is complicated to implement, but I can understand if others feel that way. My preference would be for the flag based approach but if that's deemed too complicated, then the OR approach. That said, we should discuss multiple signature evaluation at the interim Wednesday to find what the WG consensus is. -Wei
_______________________________________________ Ietf-dkim mailing list -- [email protected] To unsubscribe send an email to [email protected]
