Hello Wei,
The first is that making the handling of an unimplemented algorithm a
SHOULD has an undefined effect: the same flag produces AND at a
verifier
that implements every algorithm present, the OR outcome at one that
does
not and whatever local policy prefers elsewhere. So the signer cannot
know what its own flag will produce.
Apologies. I'm not following. Consider creating an example?
A signer places two signatures at one hop, rsa-sha256 and mldsa44 and
sets the flag to AND.
A verifier that implements both checks both. The hop passes only if both
verify, which is what the signer asked for.
A verifier that implements only rsa-sha256 follows the recommended
handling and treats the mldsa44 signature as acceptable, so the hop
passes on the rsa-sha256 signature alone, which is the OR outcome.
A third verifier implements both, finds that the mldsa44 signature fails
while the rsa-sha256 signature verifies and delivers anyway on the
ground that the newer algorithm is still being brought up in its own
deployment.
That third case is not hypothetical: cases have been reported in which
an ed25519 signature fails while the RSA signature on the same message
verifies and the message is delivered on the latter.
Three conformant verifiers, one message, one flag and only the first did
what the flag asked. The second was prevented by what it had implemented
and the third decided otherwise, and nothing in the message
distinguishes the three afterwards.
The second point is the result values. In RFC 8601 Section 2.7.1
"policy" means the message was signed but some aspect of the signature
was not acceptable to the ADMD, which is not the case of a verifier
that
cannot evaluate it. "softfail" is not a result of the dkim method at
all
and belongs to spf. The nearest existing value is "neutral",
I'm simply trying to distinguish the local policy cases that will be
defined for future
feedback reports. A sender needs to know how receivers are adopting a
new
algorithm they sign with, and needs to be distinguished from dkim2-spec
PASS and
FAIL results.
How a receiver treated an algorithm is a DKIM2 outcome, because DKIM2 is
what uses those signatures. It cannot be both separate from PASS and
FAIL and a statement about how those signatures were verified.
A sender can learn this only from a report whose outcomes come from a
standard and shared set of values, that is from a vocabulary defined in
a specification. Such a report can say, for each algorithm the sender
signed with, how many signatures the receiver received and how many fell
under each defined outcome, which for this discussion means verified,
failed or not evaluable by that receiver. The reason is then carried by
the vocabulary and the number by the report, which is how DMARC
aggregate reports already work.
If the flag is used so that a cautious sender obtains something like
composite semantics without the composite, it does not obtain it. A
composite is all or nothing: a verifier either implements it or can
check nothing of it. Two signatures with a flag are not all or nothing,
because a verifier that lacks the newer algorithm proceeds on the older
one. That is precisely the population the caution was aimed at. It is
also the behaviour that keeps interoperability at its maximum.
If the threat is as remote as you describe, the cautious sender gains
nothing from the flag. If it is not remote, the flag does not protect,
because it degrades exactly where the newer signature cannot be read.
Be careful with expectations here. RFC 8463 Section 5 says that signers
SHOULD implement Ed25519-SHA256 and that verifiers MUST implement it.
Eight years on, signed traffic carries almost no Ed25519 signatures. The
reasons given in this thread are that there was no benefit while RSA
held and that the newer algorithm failed at some sites, so senders
stayed with RSA. A requirement placed on verifiers did not produce
adoption by signers.
Regards,
Vittorio Moccia
_______________________________________________
Ietf-dkim mailing list -- [email protected]
To unsubscribe send an email to [email protected]