I share your concerns about centralization of keyserver infrastructure. Rejecting this security fix doesn't help keep keyservers decentralized, though.

Do we want to encourage multiple cleartext wire-format representations of the same secret key?

I've mentioned this interop issue (and tried to propose clarifying language for the revised standard) in the IETF OpenPGP WG mailing list.

I think rGba321b60bc3bfc29dfc6fa325dcabad4fac29f9c has nothing to do with interoperable formats -- how things are stored in ~/.gnupg/private-keys-v1.d is unrelated to the interoperable transferable secret key format specified in 4880 or its revisions.

The problem here appears to be that the "MPI" of the curve25519 secret key is not actually a standard-issue big-endian OpenPGP MPI -- it's an opaque bytestring expected to be passed to the underlying "native" implementation of x25519, in the same way that the secret key is handled for Ed25519.

investigating the subkey in python:

looks to me like you've got the byte ordering of the Curve25519 secret subkey reversed from the way that GnuPG expects it.

fwiw, gpg-agent complains that the keys don't match:

why not use gpgconf with the dirmngr component to set the keyserver option there?

Another solution to make life easier for gpgme users encountering this stuff would be if gpgme itself knows which uid is a DN and which is not, it could populate the gpgme_user_id_t.address field with content of the 1.2.840.113549.1.9.1 DN component. (or maybe gpgme_user_id_t.email, or both? as a user of gpgme, i don't really understand the difference between these fields)

fwiw, RFC 2253 is obsoleted by rfc 4514 -- which also doesn't have 1.2.840.113549.1.9.1 associated with "EMAIL", but does provide more detailed guidance for implementers of DN-to-string (and string-to-DN, to the extent that this is possible) conversions. Maybe the code should be updated to refer to the non-obsolete specification at least.

I'm reporting this because the above message renders poorly in notmuch -- notmuch gets the user ID from gmime's g_mime_certificate_get_user_id, and gmime populates that field from the uids field of a gpgme_key_t object, and gpgme pulls uid information from gpgsm --with-colons.

Attached is a proposed patch.

Attached is an even worse PKCS7 blob, that should be validatable given reliance on ca.rsa.crt, but it will be rejected by gpgsm because the PKCS#7 bundle includes ca.rsa.cross2.crt in it.

OK, i have replicated this successfully with no ed25519 involved. here's the new intermediate cert:

Which NIST test suite are you referring to? It might not cover certificate pathfinding in the face of multiple cross-signed authorities.

Apparently only one of the secret keys is actually imported: the decryption key, but not the signing key.

thanks, @werner!

Thanks for the fixes, @werner!

Other ways that gpgsm --quiet is not quiet:

I don't think the patch made elementary and ecore-x dev headers an absolute hard requirement; in particular, ./configure --disable-efl works fine to build pinentry without having these headers installed.

Thanks for the verification, @wltjr. I've pushed 19a18ba5fee049aac87b5114763095aaeb42430f to the master branch for future releases.

hm, actually, maybe the efl should be EFL in order to produce and substitute the EFL_CFLAGS and EFL_LIBS variables.

@wltjr maybe it needs ecore-x as well as elementary > 1.18 in the PKG_CHECK_MODULES line? oh, and looks like i screwed up and used > where i should have used >= sorry! fixing those would make the PKG_CHECK_MODULES line be:

I think you're saying "GnuPG will reject all subpackets marked with a critical flag unless there is a specific known semantic for *criticality* for that subpacket" Am I understanding that right? Is there a published list of criticality semantics that GnuPG is willing to accept? How do those semantics differ from standard semantics for the packet in question?

fwiw, i think a patch like this ought to work with reasonably-modern versions of autotools:

@wltjr maybe you could take a look at this?

RFC 4880 says:

See also https://gitlab.com/openpgp-wg/webkey-directory/-/issues/3 which is the same issue.

This ambiguity appears to be the cause of a recent epic (and to me, largely incomprehensible) thread on gnupg-users. It would be great to have the WKD guidance about fallback strategy be much more explicit. Any room for ambiguity here leads to different outcomes from different WKD clients, and quite a bit of confused discussion by their users.

We currently already ship:

I don't think this fix has made it into a release yet. Could we get a released version of gpgme that contains this fix?

on #debian-dpkg on IRC, Guillem Jover suggested that we might want to fix dpkg specfiles to use +self_spec: instead of *self_spec:.

I'm still not understanding what specifically should be fixed here. Sorry to be dense about it, but the range of options and configuration details that are different are pretty puzzling.

The same concern has been reported at https://bugs.debian.org/964033 -- if dirmngr is not going to follow the specification, it should at least document (and maybe warn?) about how it is divergent.

Can you characterize the failure when ipv6.disable=1 ? The straightforward failure (connect() fails with EHOSTUNREACH after a few seconds) should presumably be treated the same as if some other host happened to be offline. That should result in dirmngr failing over to the next available address for the configured keyserver, right?

I agree with you that a certificate with a lengthy expiration is not cryptographically sensible or wise, @bernhard -- i'd never want to produce such a certificate myself.

This appears to still be a problem, despite upgrading to libksba 1.4.0:

If there's no objection to this in a few days, i'll go ahead and merge it to master.

branch dkg/fix-4952 contains this fix in an easily applicable form as 0db8c768843db3e85935b972f1ed9d1b98159c46

thanks!

Thanks for following up!

To be clear: marking this ticket wontfix means (among other things) that it is the GnuPG project's upstream position that:

OK, i've asked on gnupg-devel.

In T4883#133467, @werner wrote:

That option does the same as --disable-dirmngr which in trun has the same effect as disable-crl-checks

@werner wrote:

I see no difference between the last two example stanzas that show you running ../run-verify. Are they supposed to have different output?

I'm aware of the metadata leakage risks of OCSP, and i share your concerns about them.

If CRLs or OCSP are a MUST in a given profile, and the cert chain has OCSP but no CRL, it seems like that profile should then try OCSP, rather than failing.

I think what you're saying that there is *no way* to use GPGME in offline mode to validate x.509 certificates, and this is by design. Am I understanding that right?

Thanks for the quick fix, @werner!

I thought i'd try with other certificates. I started with the one from this website. It also fails to validate unless i supply --disable-crl-checks, apparently because the immediate issuer (the Let's Encrypt CA) doesn't offer CRLs, only OCSP responders. Perhaps --disable-crl-checks should be the default, or at least if there is no CRL available there shouldn't be a failure by default:

Aha, i can get it to say f if i use --disable-crl-checks:

i didn't know that, thanks. i'm now seeing i (which i think means "invalid") in the same configuration:

For reference, here's an error message from openssl smime when it is trying to verify an e-mail message with no embedded certificate at all (despite it knowing about the relevant certificate):

"log" and "lock" are easy typo/confusions to make, @aheinecke was just trying to understand your report better, since there wasn't much information in it.

@wiktor-k, "just extend the spec" doesn't necessarily work with existing clients, which might be surprised to find unexpected packets in the signature section of an e-mail. It seems more likely to me that they'd be able to handle (meaning: ignore) an unknown subpacket (as long as it's well-formed) than to handle additional packets. But all of these surmises require testing with existing clients, of course. Has anyone done any of that testing?