[openpgp] 1PA3PC: first-party attested third-party certifications (making Key Server Prefs no-modify actionable)

[Daniel Kahn Gillmor <[e-mail]>]

Hi OpenPGP folks--

The "No-modify" flag for Key Server Preferences has never been
particularly actionable, because the keyserver has no way to tell
whether the certificate-holder intends for third-party certifications to
be redistributed or not.

After a few discussions with other implementers offlist [0], i'm
proposing a new, simple way that the certificate holder (the "first
party") can attest to some number of specific third-party
certifications, to indicate that they're redistributable.

[0] https://gitlab.com/dkg/draft-openpgp-abuse-resistant-keystore/issues/1
    and several discussions on #hagrid on IRC

The shorthand for this "first-party attestation of third-party
certifications" is 1PA3PC.

The abuse-resistant keystores draft makes clear that such an attestation
is necessary to be able to distribute third-party certifications while
retaining first-party sovereignty over the certificate.  But that draft
(up through version -04) proposes a more-complicated mechanism using
third-party confirmation signatures embedded in unhashed subpackets.

The new mechanism i'm proposing here is just a list of digests over
third-party certifications, stored in a hashed subpacket in the
self-sig.  This has two major advantages over the more complex proposal:

 * it's simpler to specify and to implement.

 * the first party can pretty easily change their mind about their
   preferred attestations for a given User ID by issuing a new self-sig;
   no need for juggling/redistributing explicit revocations of the
   attestations from the first party, which made the abuse-resistant
   keystore draft much uglier.

There are two downsides as i see it, neither of them particularly bad:

 * there is an overall space limitation in the hashed subpackets (64KiB
   tops, probably less given other overhead), so the keyholder can't
   attest to arbitrarily many third-party certifications.  However, even
   when using SHA512 (64B per attested certification), with even tighter
   constraints (like autocrypt, where the whole cert must be < 10KiB),
   there is still room for attestation of dozens of third-party
   certifications.  That's more than enough for reasonable use.

 * It encourages/requires certificate holders to issue new self-sigs
   when they want to change their attestations.  If a certificate store
   that doesn't prune superseded self-sigs encounters a rapidly-changing
   set of attestations, this might cause a bit of bloat.  But the
   certificate store can just start pruning to fix that problem.

This new proposal (diff for RFC 4880bis attached) claims subpacket
codepoint 37 for shipping these attestations.  I've also put this
proposal as a merge request here:
https://gitlab.com/openpgp-wg/rfc4880bis/merge_requests/20

1PA3PC test vector and example
------------------------------

Below is an example of this form of 1PA3PC, using ed25519.

The third-party certifier:

-----BEGIN PGP PUBLIC KEY BLOCK-----

xjMEXWYJVBYJKwYBBAHaRw8BAQdAFa8edTtqPPjjMVwac8Z+LCTbQLJv8zE7YI/8
TB8Wo7vNJkNlcnRpZmljYXRlIEF1dGhvcml0eSA8Y2FAZXhhbXBsZS5jb20+wm0E
ExYIABUFAl1l0RQCGwECFQgCF4ACGQECHgEACgkQCzDFEKWiYTfSTgEAr5HQLoxx
JLEYiEwLfixmGj5O0egfuQ8w/j+TovDvacMA/iUe9H8oyxqGKFoTn5hj0abG72mL
BWGh2++7J+BpUI8A
=W+bG
-----END PGP PUBLIC KEY BLOCK-----

The end user certificate, with the third-party certification and a
self-sig that attests to it:

-----BEGIN PGP PUBLIC KEY BLOCK-----

xjMEXWYJVBYJKwYBBAHaRw8BAQdA3RPxU734ifWKfZMLChBCYYCD3Pq6mV1RFyhh
H4vSq77NHFRlc3QgVXNlciA8dGVzdEBleGFtcGxlLmNvbT7CXgQQFggABgUCXWXR
HgAKCRALMMUQpaJhN3V/AQD8tnvWlpuXEjPDX4jxJ9COhobYS3UG8Id4i0Mhe/4c
wAD8DehW+aDn7+Nv0qPJrUuyDeODHj7/MCvvEaxltrgjkQ/CjwQTFggANwUCXWXR
KAIbAQIVCAIXgAIZAQIeASElZykynZ95I3ov66EDuDZE1g6By2V0gE0Z0RlkE1DT
CsQACgkQ1NdAbG04n/QVxwEAgZ812nc7lqISocouns7Z7Zi7LHj8RyI9JrpmmDio
fXkA+wRUUxgsQ63nyWfOJDvMl5Z57IKQOvlWJrtmSZ9ORGwG
=ApYY
-----END PGP PUBLIC KEY BLOCK-----

I'd be happy to hear any feedback, both about the proposed patch to the
spec, and about the test vector i've included here.

      --dkg