Oopsie, I just saw Laolu's footnote #4 about ignoring isogeny crypto. Guess I should learn to read better :P

regards,

conduition

On Friday, May 8th, 2026 at 8:21 PM, conduition <condu...@proton.me> wrote:

Hi all,

I'm not well-versed in the P2P network transport protocol or BIP324, so I'm not well-qualified to give feedback on the details of this idea. But I did want to chime in on this statement:

One thing worth noting is that AFAICT, so far in the NIST PQC world [4], there is no known non-interactive key exchange protocol like we enjoy today with ECDH. IIUC, the reason is that lattice based schemes derived from the LWE [3] problem, whose security is predicated on using "noise" to hide a secret value. For these cryptosystems, usually a type of "hint" is sent to make everything work out nicely like in ECDH. However, in the stricter non-interactive setting (no messages sent), this doesn't map cleanly. 

It's important to emphasize this only considers the NIST-standardized KEMs. If we zoom out to the broader ecosystem of PQ PKE candidates, there are several options for non-interactive key exchange systems that'd be a drop-in replacement for ECDH.

For instance, oriented isogeny-based systems like CSIDH [1] [2] permit this kind of construction. Both parties publish a short (64 to 128 byte) pubkey and can perform key exchange as soon as they've seen the peer's pubkey, no additional messages required. The down side of CSIDH is that it's quite slow, so it is most useful when pubkeys are static or otherwise don't change much. There has been a lot of work done with new faster schemes [3] or speeding up CSIDH with better implementations [4] but still key exchange can take a good few dozen milliseconds. 

In general, any post-quantum-secure commutative group action scheme allows non-interactive key exchange. CSIDH is just one such example, and I'm sure there are and will be others.

Being unversed in BIP324 as yet, I'm not sure how crucial this non-interactivity property is to the protocol. If it's not a big deal, then I'd gladly toss my hat in for lattices and a hybridized ML-KEM construction, given so much of the internet is already migrating to this. It makes sense to follow standards if we can. Going full-TLS would probably be overkill IMO - It is designed for a very different (centralized) PKI architecture and would buy us a ticket for a train we probably don't want to ride.

Maybe there's a more applicable standard, a la Noise/Wireguard? For a possible implementation reference, see [5]. Otherwise rolling our own standard seems like the way to go, especially if we can do so in a way that is reusable for other use-cases beyond Bitcoin.

Also, if we go with a hybrid scheme, we should have clear migration paths to transition to either pure PQC (if a CRQC appears and breaks ECDH, we might as well discard it), or back to classical ECDH (if CRQCs turn out to be impossible).

regards,

conduition

[1]: https://csidh.isogeny.org/

[2]: https://eprint.iacr.org/2018/383

[3]: https://eprint.iacr.org/2025/1098.pdf

[4]: https://ctidh.isogeny.org/ctidh-20210513.pdf

[5]: https://github.com/jmlepisto/clatter

On Thursday, May 7th, 2026 at 1:45 AM, Jonas Schnelli <jonas.s...@gmail.com> wrote:

Thanks for writing this up Laolu.

I think option 1 (classical-then-PQ-upgrade) is probably the right path, mostly because it keeps the byte-0 pseudorandomness property without needing Kemeleon or any new obfuscation primitive.

If the ML-KEM exchange happens inside the already-established v2 ChaCha20Poly1305 channel, then to a present-day classical observer the bytes should still look random,... the inner PQ handshake is just more ciphertext. A future QC adversary doing harvest-now-decrypt-later would break the outer ECDH eventually, but I think they'd then still have to break ML-KEM-768 to get the v3 transport keys, which is kind of the whole point. So we'd probably get hybrid security against the QC threat and also keep the property that today's wire bytes are indistinguishable from random.

That said, I'm not sure how far we should really take the pseudorandomness argument,... traffic shape (packet sizes, timings, query/response patterns) probably already reveals quite a bit about what's going on, so the byte-content randomness is only one part of the picture.

The extra round trip is probably fine given how long Bitcoin P2P connections live. The DoS angle you flagged might also be smaller in option 1, since the responder still commits after 64 bytes and not after a 1184-byte ML-KEM key,... though I haven't thought hard about wether there are other DoS vectors the inner upgrade introduces.

On Ethan's TLS 1.3 suggestionm,... I don't think it really fits. Apart from the dependency cost (which we deliberately kept low in BIP 324), TLS has it's own fingerprint, which would probably undo the censorship-resistance angle. And it bundles authentication with encryption, which we explicitly decoupled.

One thing worth looking at: OpenSSH (whose chacha20-poly1305 construction we drew from originally) shipped mlkem768x25519-sha256 as default in 10.0 last year, and they just concatenate-and-hash the two shared secrets. Their threat model doesn't care about pseudorandomness so they can send ML-KEM material in the clear, but the combiner shape is probably a reasonable reference for ours.

/jonas

On May 6, 2026, at 12:15 PM, Olaoluwa Osuntokun <lao...@gmail.com> wrote:

Hi Ethan, 

That's a great question. 

First, I don't speak for Bitcoin Core by any means (btcd has also implemented
BIP 324 FWIW). Based on past observed behavior, they typically prefer to keep
dependencies slim. Many years ago there was a concerted push to remove openssl
as a dependency from the project. So I would imagine the idea of rolling out
full blown TLS 1.3 might encounter some resistance.

In terms of cryptography, BIP 324 as defined uses secp256k1. TLS 1.3 as
specified doesn't support secp256k1 within the set of supported cipher suites.

If ensuring that BIP 324 continues to implement an oblivious KEM is a key
requirement, then TLS 1.3 doesn't fit the bill.

Regarding a hybrid PQ KEM, there exists an IETF to add a new key agreement
suite to TLS 1.3: https://datatracker.ietf.org/doc/draft-ietf-tls-ecdhe-mlkem/.
Only secp256+384(r1) and x25519 are supported as elliptic curves in this draft.

One additional aspect is that today BIP 324 doesn't implement authentication at
all, you only get confidentiality. TLS 1.3 would mean introducing certificates
in some fashion, thereby coupling concerns from the original PoV of Bip 324.

BIP 324 also includes as section in the BIP detailing the rationale of BIP 324
over a more general purpose protocol (mentions some of the points above):
https://github.com/bitcoin/bips/blob/master/bip-0324.mediawiki#:~:text=Why%20not%20use%20a%20general%2Dpurpose%20transport%20encryption%20protocol%3F.


-- Laolu

On Tue, May 5, 2026 at 2:18 PM Ethan Heilman <eth...@gmail.com> wrote:

Thanks Laolu for thinking through making PQ BIP 324 and writing this up.

Reading through what you wrote made me what wonder, why not use this as an opportunity to move to TLS 1.3?

What's the case against using TLS 1.3 for PQ P2P connection encryption? Is there some functionality that TLS 1.3 is lacking that we really want?  Is the case against solely to not have TLS 1.3 as a complex dependency in bitcoin-core?

The advantages of TLS 1.3:

1. Make Bitcoin P2P connections blend in with all the other TLS connections. This isn't strong privacy, you can distinguish TLS encrypted Bitcoin traffic via timing and size, but it reduces accidents where a firewall sees an unknown protocol and blocks it.

2. Use of QUIC for faster relay, oblivious HTTP and QUIC tunnels-in-tunnels for private relay and similar protocols.

3. Lots of eyeballs on TLS 1.3, we don't need to build or maintain it.

On Tue, May 5, 2026, 00:41 Olaoluwa Osuntokun <lao...@gmail.com> wrote:

Hi y'all, 

In case you weren't already tired of all the recent dev list chatter re post
quantum cryptography, here's another!

When the topic of Bitcoin transitioning to a post quantum world is brought up,
the discussion typically focuses on the consensus layer re swapping out
vulnerable signature schemes. However, the consensus layer isn't the only area
of Bitcoin that relies in cryptography that would be broken in the face of a
powerful quantum computer! That's right, I'm talking about BIP 324, the peer to
peer encryption BIP for Bitcoin.

Like everything else on the Internet today, BIP 324 uses ECDH to allow two
connecting peers to derive a shared secret known only to them, which is then
used to encrypt all traffic between them. As ECDH relies on Elliptic Curve
cryptography, a future quantum computer would be able to eavesdrop on a p2p
handshake transcript, then derive the underlying private keys to the ephemeral
ECDH public key, permitting it to decrypt all traffic. It's actually worse than
that, as today adversaries can collect all encrypted p2p Bitcoin traffic, with
the hope of being able to decrypt it all at a future date. This is commonly
referred to as the: "harvest, decrypt later" (HNDL) strategy [11].

Compared to a consensus change, which requires widespread market agreement, and
coordination to achieve, upgrading BIP 324 to be post quantum resistant is a
much lower hanging fruit worthy of pursing immediately.

Last week I starting thinking a bit about this topic, brushing up on the latest
literature/techniques, and stumbled onto a few key design questions. The goal
of this post isn't to propose a new concrete p2p encryption BIP, instead I want
to start discussion on the various design tradeoffs that came up as I was
researching this p2p encryption transition.

## PQ BIP 324 Design Questions

1. Do we want to pursue a hybrid KEM (key encapsulation mechanism), or go with
   a pure PQ KEM?

2. Is it still a key requirement that the initial handshake be
   indistinguishable from a random byte string?

   2a. If yes to the above, then should we go with classical-then-pq-upgrade,
   or a one shot hybrid oblivious KEM.


## A Brief Intro to KEMs + ML-KEM

First, let's introduce the new primitive we have to work with: ML-KEM
(Module-Lattice-Based Key-Encapsulation Mechanism) [1][2]. As it says on the
tin, ML-KEM is a lattice based Key-Encapsulation Mechanism. The phrase KEM
might sound unfamiliar with those comfortable with ECDH, but ECDH is actually a
KEM itself.

A KEM has 3 algorithms:
  * KeyGen() -> {sk, pk}
     * Generates a public/private secret key pair

  * Encaps(pub) -> {secret, capsule}
     * Generates a new secret value, and a "capsule", which only the holder of
       pub can use to obtain the secret value.

  * Decaps(priv, capsule) -> secret
     * Uses the private key to extract the secret from the capsule


If you squint a bit, then you'll see that ECDH is a KEM, and a rather elegant
one at that:
  * KeyGen() -> {k, k*G}
      * Normal EC key generation. 

  * Encaps(pub) -> {capsule = x*G, secret = pub*x}
      * The core ECDH routine. The ephemeral public key is actually the
        "capsule". The resulting secret is the ECDH output with the remote
        party's KEM public key and the local secret.

  * Decaps(priv, capsule) -> secret = priv * capsule
      * The receiver completes the key exchange using the ephemeral public key
        and their own private key.

ECIES is another flavor of EC based KEM.

One thing worth noting is that AFAICT, so far in the NIST PQC world [4], there is
no known non-interactive key exchange protocol like we enjoy today with ECDH.
IIUC, the reason is that lattice based schemes derived from the LWE [3]
problem, whose security is predicated on using "noise" to hide a secret value.
For these cryptosystems, usually a type of "hint" is sent to make everything
work out nicely like in ECDH. However, in the stricter non-interactive setting
(no messages sent), this doesn't map cleanly.

As a result, ML-KEM looks more like a hybrid encryption protocol (Alice
encrypts a shared secret to bob using asymmetric lattice crypto).

## To Hybrid KEM, Or Not to Hybrid KEM

This brings us to our first design question....

Should we use a hybrid KEM or a pure post quantum one? 

A hybrid KEM would keep the existing ECDH, _also_ do ML-KEM, then securely
combine (there's some subtlety there, see [6][7]) the resulting in a
final secret value for encryption. A hybrid KEM is attractive as an encryption
channel derived from such a KEM is secure if _any_ of the combined schemes are
secure. This permits schemes to hedge a bit, as hey, maybe the PQ stuff is
actually broken in the future but ECDH isn't. If it's the other way around,
then your encryption scheme is still secure.

### Pure ML-KEM P2P Encrypted Handshake

If we opt to not use a hybrid scheme, then the Elligator layer can be dropped
all together. Instead, the 1.1 KB (ML-KEM-768) encapsulation keys are sent,
keeping the trailing garbage+terminator in tact. 

The initial handshake would look something like: 
 * Alice -> Bob: alice_encaps || initiator_garbage
    * Alice derives an encapsulation key, and sends it to Bob.

 * Bob -> Alice: ml_kem_capsule || responder_garbage || responder_garbage_terminator || first_encrypted_packet
   * Bob uses Alice's encapsulation key to encapsulate a random secret, and
     sends it over to Alice. He can also encrypt the first message at this
     point.

 * Alice -> Bob: initiator_garbage_terminator || first_encrypted_packet
   * Alice de-encapsulates the shared secret, and can now also start to encrypt
     messages.

We'd then replace `v2_ecdh` with something like a `v3_mlkem` that derives the
final shared secret based on the sent/received transcript up until that point:
  * `sha256_tagged("bip324_ml_kem", ml_kem_secret, alice_encaps, ml_kem_capsule)`

### Hybrid ML-KEM P2P Encrypted Handshake

If we want to use a hybrid combiner, then along side the normal ellswift keys,
the ML-KEM-768 encap key is also sent:

 * Alice -> Bob: ellswift_alice || alice_encaps || initiator_garbage
 * Bob -> Alice: ellswift_bob || ml_kem_capsule || responder_garbage || responder_garbage_terminator || first_encrypted_packet
 * Alice -> Bob: initiator_garbage_terminator || first_encrypted_packet

Then following guidelines of [7], we'd then replace `v2_ecdh` with something
like `v3_hybrid_shared_secret`:
  * `sha256_tagged("bip324_ellswift_xonly_ecdh_mlkem_768", ml_kem_ss, ecdh_point_x32, alice_encaps, ml_kem_capsule, ellswift_alice, ellswift_bob)`

## PQ/Hybrid Obfuscated KEMs

At this point, those that are familiar with BIP 324 will recognize that both
the pure PQ and hybrid versions renders the ElligatorSwift usage pretty much
useless. ElligatorSwift encodes a 32-byte public key as a 64-byte value which
is indistinguishable from a uniformly distributed bitstream. In a bubble, this
means that the initial BIP 324 handshake to a 3rd party observer just looks
like random bytes. However, with the introduction of ML-KEM, the ML-KEM
encapsulation key is sent in plaintext over the wire. An ML-KEM key has
identifiable structure, as it's a giant vector of polynomial coefficients mod
3329, which is easily recognizable over the wire.

Luckily, there's an ML-KEM analogue to ElligatorSwift, called Kemeleon
[8][9][10]! In a similar fashion to ElligatorSwift, it takes an ML-KEM public
key, then encodes it as one giant integer, utilizing rejection sampling.
Kemeleon applies this mapping both to the encapsulation keys, and also the
capsule ciphertext that encrypts the shared secrets. The ML-KEM keys end up
being a bit smaller, while the ciphertexts map to a larger value. Another
tradeoff is that the Kemeleon key generation is ~3x slower than normal ML-KEM
generation.

One thing to note here is that Kemeleon's "looks random" property isn't quite
on the same footing as ElligatorSwift's. ElligatorSwift is statistically
indistinguishable from random, since every 512-bit string is a valid encoding.
Kemeleon's indistinguishability is computational, resting on a Module-LWE
style assumption. So if you naively concatenate an ElligatorSwift key and a
Kemeleon key, the pair is only as obfuscated as the weakest visible half. This
asymmetry is what motivates the OEINC construction discussed below.

This brings us to our second design question....

Do we still want to ensure that the BIP-324 handshake looks identical to a
pseudorandom bytestream from the very first message?

Assuming yes, then AFAICT, we have two classes of options here: 
  1. Retain the existing BIP-324 outer ElligatorSwift handshake, but use ML-KEM
     within that initial encrypted transport to upgrade to a PQ shared secret.

  2. Use the Outer Encrypts Inner Nested Combiner (OEINC - "OINK") combiner
     from [8].

  3. Attempt to adapt Drivel from [8] into the Bitcoin p2p setting.

### Classical Encrypted Channel Upgrades to PQ

With the first option, we simply use one KEM right after the other. So BIP 324
v2 would be mostly unchanged, then we _upgrade_ to BIP 324 v3 within v2. 

A sketch of this would be something like:
  * Phase 0: normal BIP 324 handshake
  * Phase 1: negotiation of PQ KEM scheme over the encrypted handshake
     * Can be optional, if we just pick a set PQ KEM scheme.
     * Before this point, no Bitcoin p2p message should be sent, as the channel
       isn't PQC protected yet.
  * Phase 2: do normal ML-KEM within the ElligatorSwift derived encrypted
    transport
     1. Alice sends the encapsulation key
     2. Bob derives a secrets, encrypts it using the encapsulation key
     3. Both sides then derive a PQ shared secret, ss_PQ
  * Phase 3: both sides use a hybrid combiner like sketched out above to derive
    a new set of transport keys
  * Phase 4: both sides rekey, switching over to a new the transport keys

The upside of this option is that the outer part of BIP 324 remains unchanged,
then with another round trip, we're able to upgrade the encryption keys to PQ
hybrid security. The downside is that the very first messages sent aren't PQ
from the start, but a PQ adversary wouldn't be able to decrypt the actual
Bitcoin p2p messages (as we wait to send those until the upgrade). The
handshake still looks like just random bytes.

### Outer Encrypts Inner Nested Combiner

For the second option, [8] (with talk video [9] and slides [10]) describes an
OEINC scheme where the outer KEM
encrypts the inner KEM, wherein the KEM ciphertext of an inner KEM is encrypted
using a shared secret derived from the outer KEM. The two KEM ciphertexts and
the two derived keys are then used alongside a hybrid combiner to derive a
final shared secret. 

Unlike the classical-then-pq-upgrade that establishes a classical channel, then
uses that to upgrade to pq channel, OEINC is a special hybrid combiner that
achieves a similar output but in one swoop. It defines a special KEM, which can
then be used as the KEM in the very first handshake I sketched out.

A sketch of this KEM looks something like:
  * Setup:
    * The outer KEM is BIP 324's ElligatorSwift-encoded secp256k1 DHKEM.
       * It serves as the outer KEM because its on-wire encoding is
         statistically indistinguishable from random.
    * The inner KEM is ML-Kemeleon.

  * KeyGen():
    * (kem_secret_outer, kem_pubkey_outer) = outKEM.Gen()
    * (kem_secret_inner, kem_pubkey_inner) = inKEM.Gen()
    * combined_pubkey = (kem_pubkey_outer, kem_pubkey_inner)
    * combined_secret = (kem_secret_outer, kem_secret_inner)

  * Encaps(combined_pubkey):
    * (shared_secret_outer, capsule_outer) = outKEM.Encap(kem_pubkey_outer)
    * (encrypt_key_1, encrypt_key_2) = KDF(shared_secret_outer)
    * (shared_secret_inner, capsule_inner) = inKEM.Encap(kem_pubkey_inner)
    * encrypted_capsule_inner = encrypt(encrypt_key_1, capsule_inner)
    * combined_capsule = capsule_outer || encrypted_capsule_inner
    * combined_shared_secret = combine(encrypt_key_2, shared_secret_inner, combined_capsule)

  * Decaps(combined_secret, combined_capsule):
    * (capsule_outer, encrypted_capsule_inner) = combined_capsule
    * shared_secret_outer = outKEM.Decaps(kem_secret_outer, capsule_outer)
    * (encrypt_key_1, encrypt_key_2) = KDF(shared_secret_outer)
    * capsule_inner = decrypt(encrypt_key_1, encrypted_capsule_inner)
    * shared_secret_inner = inKEM.Decaps(kem_secret_inner, capsule_inner)
    * combined_shared_secret = combine(encrypt_key_2, shared_secret_inner, combined_capsule)


This is done over just sending the two encapsulated secrets plainly as I
outlined above in order to achieve a stronger security notion. The issue with
this though is that though ciphertext uniformity (the encapsulated secrets) is
achieved, the two public keys sent are randomly looking, but not in a uniform
manner. In practice, this might not really matter much AFAICT (a theoretical
adversary would be able to distinguish the Elligator half from the Kemeleon
half).

### Drivel: PQ-Obfuscated Authentication

The biggest issue with Drivel as a fit for BIP 324 is that it expects the
initiator to already know a long term static public key for the responder. In
the case of BIP 324, only ephemeral keys are exchanged, so there's no long
term public keys known to either side.

To get around this, we could extend BIP 155 (or make a new one likely, given
size limits) to include a signed OKEM key. However then that would introduce
authentication into the combined set, which explicitly wasn't a design goal
of BIP 324.

With that caveat in mind, here's the construction itself. Drivel [8] combines
the OEINC scheme with another layer that out-of-the-box assumes an asymmetric
protocol within a set client and server. The client uses an existing OEINC
KEM public key published by the server to then encrypt a fresh new ephemeral
KEM.

----- 

So there we have it. Before drafting a concrete v3 transport, we need to
decide if we want a hybrid KEM, or are fine with a pure PQ KEM. Then we need to
decide if we want to attempt to maintain the current quality where the p2p
handshake transcript is indistinguishable from random. If yes, then that forces
another series of decisions re how to construct/compose an oblivious KEM from
available primitives.

At a glance, the route of classical-then-pq-upgrade seems to be the simplest.
BIP 324 stays as is, then we run ML-KEM within that. The ML-KEM keys are
encrypted, so there's no need to sprinkle in the layer of Kemeleon.

If we want a nice combined protocol, then we should investigate the OEINC
route. It's more data to send as part of the initial handshake, but we still
keep ElligatorSwift and use that as the outer KEM.

If for some reason we're concerned with a future adversary gaining a
distinguisher for Kemeleon, then maybe we need to bite the bullet and also
roll out a full blown PQ authentication protocol along side everything.

One thing worth flagging for any of the byte-0 designs (where PQ material is
sent in the clear on the very first flight, like the hybrid and OEINC sketches
above): ML-KEM-768 makes the responder do real work before it can decide if a
connection is even legit. Today, the responder only needs the first 64 bytes
of an ElligatorSwift share before it can derive the shared secret. With
ML-KEM-768, the responder has to read and validate a 1184 byte encapsulation
key before running Encaps, and FIPS 203 mandates input checks on every Encaps
and Decaps. In a permissionless P2P network, that's a meaningful change in
inbound DoS surface, and probably calls for stricter handshake byte limits,
tighter timeouts, and possibly some form of stateless cookie/puzzle if
handshake floods become a real problem. The classical-then-pq-upgrade path
sidesteps most of this since the PQ material only shows up after the v2
channel is up.

With all that said, after the above design decisions are addressed, there
aren't too many concrete blockers here w.r.t rolling this out. Of course the
development (eg: selecting/creating a library for ML-KEM and maybe
ML-Kemeleon), and upgrade will take some time. But unlike the consensus
layer, p2p encryption doesn't require the widespread market agreement that an
actual soft fork does. BIP 324 is a much shorter walk to PQ than the consensus
layer, and serves as a sort of PQ warm up before the bigger soft fork is
tackled. 


-- Laolu

[1]: https://en.wikipedia.org/wiki/ML-KEM
[2]: https://csrc.nist.gov/pubs/fips/203/final
[3]: https://en.wikipedia.org/wiki/Learning_with_errors
[4]: This statement ignores Isogeny based crypto, and also SWOOSH [5] as it requires 200 KB pubkeys
[5]: https://eprint.iacr.org/2023/271
[6]: https://eprint.iacr.org/2018/024
[7]: https://eprint.iacr.org/2020/1364
[8]: https://eprint.iacr.org/2024/1086
[9]: https://www.youtube.com/watch?v=CvFCYUq5rGg
[10]: https://csrc.nist.gov/csrc/media/Presentations/2025/kemeleon/images-media/kemeleon.pdf
[11]: https://en.wikipedia.org/wiki/Harvest_now,_decrypt_later

-- 
You received this message because you are subscribed to the Google Groups "Bitcoin Development Mailing List" group.
To unsubscribe from this group and stop receiving emails from it, send an email to bitcoindev+...@googlegroups.com.
To view this discussion visit https://groups.google.com/d/msgid/bitcoindev/CAO3Pvs9U3prZJiDs0Ns7LSA07R8hM-GQou_FcTZZz-JUQpUYHw%40mail.gmail.com.

-- 
You received this message because you are subscribed to the Google Groups "Bitcoin Development Mailing List" group.
To unsubscribe from this group and stop receiving emails from it, send an email to bitcoindev+...@googlegroups.com.
To view this discussion visit https://groups.google.com/d/msgid/bitcoindev/CAO3Pvs8i%3DpLP30nRh_iyjSRJXne19wNezmQmo%3DJAk8%2BE4uJPhg%40mail.gmail.com.

--
You received this message because you are subscribed to the Google Groups "Bitcoin Development Mailing List" group.
To unsubscribe from this group and stop receiving emails from it, send an email to bitcoindev+...@googlegroups.com.
To view this discussion visit https://groups.google.com/d/msgid/bitcoindev/547ADFCA-2BB1-4E16-A26A-C92262EBBD84%40gmail.com.