The PGP Problem

Matrix works, and has fairly easy-to-use clients (and I've switched a few non-technical folks without too much trouble). It's also an open protocol, not just an app.

The main criticism (and I'm preemptively responding to tptacek here) is that they haven't yet made E2EE the default (though this should happen in a few weeks now that cross-signing appears to be done). I also think the key backup system should be much better explained -- there is a usability bug open for it and I've posted some suggestions.

Magic Wormhole is neat, but the happy path is to use a third-party rendezvous server, which is susceptible to traffic analysis (I also wish that the codes had more than 16 bits of entropy, but that is part cargo-culting on my part).

Signal is also vulnerable to server-side traffic analysis, and is strangely keen on both demanding a selector (a phone number) for identity and on trusting Intel's 'secure' enclave (I strongly suspect that it's suborned).

One thing I do like about PGP is that it has been around awhile: I can still decrypt my old files & verify old signatures just fine, something I don't trust the flavour of the month to do.

I think that rather than a Swiss Army knife tool & protocol like PGP, we should have a suite of tools built around a common core, probably NaCL. That way we can have compatibility when we need it, but also aren't cramming square pegs into round holes.

Finally, the Web of Trust was a bright, shiny idea — and wrong. But Trust on First Use is also pretty bad, as is the CA ecosystem. We need something else, something decentralised and also secure — and I don't think that's impossible. I've had a few ideas, but they haven't panned out. Maybe someday.

I did some "OpenPGP Best Practices" work for a client recently. They don't have a choice, because a third party requires it. The goal was to make sure it was as safe as possible. One thing that struck me is that I have a simplified mental model for the PGP crypto, and reality is way weirder than that. The blog post says it's CFB, and in a sense that's right, but it's the weirdest bizarro variant of CFB you've ever seen.

In CFB mode, for the first block, you take an IV, encrypt it, XOR it with plaintext. Second block: you encrypt the first ciphertext block, encrypt that, XOR with second plaintext block, and so on. It feels sorta halfway between CBC and CTR.

Here's the process in OpenPGP, straight from the spec because I can't repeat this without being convinced I'm having a stroke:

   1.   The feedback register (FR) is set to the IV, which is all zeros.

   2.   FR is encrypted to produce FRE (FR Encrypted).  This is the
        encryption of an all-zero value.

   3.   FRE is xored with the first BS octets of random data prefixed to
        the plaintext to produce C[1] through C[BS], the first BS octets
        of ciphertext.

   4.   FR is loaded with C[1] through C[BS].

   5.   FR is encrypted to produce FRE, the encryption of the first BS
        octets of ciphertext.

   6.   The left two octets of FRE get xored with the next two octets of
        data that were prefixed to the plaintext.  This produces C[BS+1]
        and C[BS+2], the next two octets of ciphertext.

   7.   (The resynchronization step) FR is loaded with C[3] through
        C[BS+2].

   8.   FRE is xored with the first BS octets of the given plaintext,
        now that we have finished encrypting the BS+2 octets of prefixed
        data.  This produces C[BS+3] through C[BS+(BS+2)], the next BS
        octets of ciphertext.

   9.   FR is encrypted to produce FRE.

   10.  FR is loaded with C[BS+3] to C[BS + (BS+2)] (which is C11-C18
        for an 8-octet block).

   11.  FR is encrypted to produce FRE.

   12.  FRE is xored with the next BS octets of plaintext, to produce
        the next BS octets of ciphertext.  These are loaded into FR, and
        the process is repeated until the plaintext is used up.

Yeah so CFB except your IV isn't your IV and randomly do something with two bytes as... an... authenticator? And then everything after that is off by two? This isn't the only case where OpenPGP isn't just old, it's old and bizarre. I don't have a high opinion of PGP to begin with, but even my mental model is too charitable.

(Disclaimer: I'm a Latacora partner, didn't write this blog post but did contribute indirectly to it.)

Re: magic-wormhole's 16 bits: I don't think you should be worried about that, because SPAKE2 will give you proof positive if the attacker attempts to guess. Are you saying 2*-16 success isn't good enough?

Isn't the code length also selectable?