[bitcoin-dev] RBF Pinning with Counterparties and Competing Interest

[Matt Corallo <lf-lists@mattcorallo.com>]

[Hi bitcoin-dev, in lightning-land we recently discovered some quite frustrating issues which I thought may merit
broader discussion]

While reviewing the new anchor outputs spec [1] last week, I discovered it introduced a rather nasty ability for a user
to use RBF Pinning to steal in-flight HTLCs which are being enforced on-chain. Sadly, Antoine pointed out that this is
an issue in today's light as well, though see [2] for qualifications. After some back-and-forth with a few other
lightning folks, it seems clear that there is no easy+sane fix (and the practicality of exploitation today seems
incredibly low), so soliciting ideas publicly may be the best step forward.

I've included lots of background for those who aren't super comfortable with lightning's current design, but if you
already know it well, you can skip at least background 1 & 2.

Background - Lightning's Transactions (you can skip this)
=====================================

As many of you likely know, lightning today does all its update mechanics through:
 a) a 2-of-2 multisig output, locking in the channel,
 b) a "commitment transaction", which spends that output: i) back to its owners, ii) to "HTLC outputs",
 c) HTLC transactions which spend the relevant commitment transaction HTLC outputs.

This somewhat awkward third layer of transactions is required to allow HTLC timeouts to be significantly lower than the
time window during which a counterparty may be punished for broadcasting a revoked state. That is to say, you want to
"lock-in" the resolution of an HTLC output (ie by providing the hash lock preimage on-chain) by a fixed block height
(likely a few hours from the HTLC creation), but the punishment mechanism needs to occur based on a sequence height
(possibly a day or more after transaction broadcast).

As Bitcoin has no covanents, this must occur using pre-signed transactions - namely "HTLC-Success" and "HTLC-Timeout"
transactions, which finalize the resolution of an HTLC, but have a sequence-lock for some time during which the funds
may be taken if they had previously been revoked. To avoid needless delays, if the counterparty which did *not*
broadcast the commitment transaction wishes to claim the HTLC value, they may do so immediately (as there is no reason
to punish the non-broadcaster for having *not* broadcasted a revoked state). Thus, we have four possible HTLC
resolutions depending on the combination of which side broadcast the HTLC and which side sent the HTLC (ie who can claim
it vs who can claim it after time-out):

 1) pre-signed HTLC-Success transaction, providing the preimage in the witness and sent to an output which is sequence-
    locked for some time to provide the non-broadcasting side the opportunity to take the funds,
 2) pre-signed HTLC-Timeout transaction, time-locked to N, providing no preimage, but with a similar sequence lock and
    output as above,
 3) non-pre-signed HTLC claim, providing the preimage in the witness and unencumbered by the broadcaster's signature,
 4) non-pre-signed HTLC timeout, OP_CLTV to N, and similarly unencumbered.

Background 2 - RBF Pinning (you can skip this)
==========================

Bitcoin Core's general policy on RBF transactions is that if a counterparty (either to the transaction, eg in lightning,
or not, eg a P2P node which sees the transaction early) can modify a transaction, especially if they can add an input or
output, they can prevent it from confirming in a world where there exists a mempool (ie in a world where Bitcoin works).
While this is somewhat unintuitive, there are any number of good anti-DoS reasons for this, eg:
 * (ok, this is a bad reason, but) a child transaction could be marked 'non-RBF', which would mean allowing the parent
   be RBF'd would violate the assumptions those who look at the RBF opt-in marking make,
 * a parent may be very large, but low feerate - this requires the RBF attempt to "pay for its own relay" and include a
   large absolute fee just to get into the mempool,
 * one of the various package size limits is at its maximum, and depending on the structure of the package the
   computational complexity of calculation evictions may be more than we want to do for a given transaction.

Background 3 - "The RBF Carve-Out" (you can skip this)
==================================

In today's lightning, we have a negotiation of what we expect the future feerate to be when one party goes to close the
channel. All the pre-signed transactions above are constructed with this fee-rate in mind, and, given they are all
pre-signed, adding additional fee to them is not generally an option. This is obviously a very maddening prediction
game, especially when the security consequences for negotiating a value which is wrong may allow your counterparty to
broadcast and time out HTLCs which you otherwise have the preimage for. To remove this quirk, we came up with an idea a
year or two back now called "anchor outputs" (aka the RBF carve-out for those in Bitcoin-land) - a neat trick to allow
both counterparties to add fees to a transaction which is being broadcast without getting into the quagmire that is RBF
pinning. Specifically, we added a rule to Bitcoin Core which allows for transactions which have a narrow structure to be
CPFP'd trivially by either counterparty, irrespective of what the other counterparty does! In order to meet this
structure, the commitment transaction (b) must have two (potentially-)additional outputs, each which only one side can
spend, and every other output must have a CSV lock associated with it. This is great and there is (finally) movement to
deploy this.

RBF Pinning HTLC Transactions (aka "Oh, wait, I can steal funds, how, now?")
=============================

You'll note that in the discussion of RBF pinning we were pretty broad, and that that discussion seems to in fact cover
our HTLC outputs, at least when spent via (3) or (4). It does, and in fact this is a pretty severe issue in today's
lightning protocol [2]. A lightning counterparty (C, who received the HTLC from B, who received it from A) today could,
if B broadcasts the commitment transaction, spend an HTLC using the preimage with a low-fee, RBF-disabled transaction.
After a few blocks, A could claim the HTLC from B via the timeout mechanism, and then after a few days, C could get the
HTLC-claiming transaction mined via some out-of-band agreement with a small miner. This leaves B short the HTLC value.

You'll note that B would be just fine if they had a way to safely monitor the global mempool, and while this seems like
a prudent mitigation for lightning implementations to deploy today, it is itself a quagmire of complexity, especially
when you consider differences in relay policy during an upgrade cycle and how those may effect propagation through the
P2P network. Further, this is a really obnoxious assumption to hoist onto lightning nodes - having an active full node
with an in-sync mempool is a lot more CPU, bandwidth, and complexity than most lightning users were expecting to face.

It seems highly likely we could come up with some kind of variant of of the RBF Carve-Out to solve this problem, though
probably much more specific to this particular transaction structure - imagine a rule which allowed B to RBF C's low-fee
HTLC claim transaction, without ever seeing it. This could be accomplished by locking down the transaction types in (3)
and (4) by pre-signing them (just, like (1) and (2)) and then using some kind of policy rule to allow only the addition
of additional confirmed inputs and one (small) output. This would mean that B knows that either C's transaction has high
fee, or B's reasonably-higher-fee transaction will meet the RBF rules and replace C's maliciousness.

While the original RBF Carve-Out was a little awkward, its structure was sufficiently generic that other off-chain
protocols could reasonably (need to) take advantage of it, however a rule to address this issue seems like it would need
to be highly tailored to lightning, which doesn't seem acceptable (there appears to be a way to shoehorn the existing
carve-out, but it results in a ton of extra on-chain volume).

Strategies involving full-RBF for transactions not at the top of the mempool, (slow-)full-mempool-sync allowing
mempool-total-fee decreases and relaxations of the RBF rules would be welcome, but without a ton of legwork to include
things like package relay I'm not convinced they would suffice. This of course doesn't even account for the possibility
of similar issues given rely policy differences.

PS For Lightning-Dev Folks (aka "An Alternative Anchor Proposal")
==========================

Given the anchor outputs proposal seeks to expand lightning's security in a world where Bitcoin is running at
steady-state and the mempool reliably has transactions in it, not fixing this issue seems to render the whole exercise
somewhat useless (not to mention that the current design makes this attack more obvious and provides several alternative
paths to exploitation).

An alternative, albeit not ideal anchor outputs proposal is as follows:

 * Instead of making the HTLC output spending more free-form with SIGHASH_ANYONECAN_PAY|SIGHASH_SINGLE, we clearly need
   to go the other direction - all HTLC output spends need to be pre-signed.
 * Sadly, and this really hurts from an on-chain-tx-size perspective, we have to include anchor outputs in the HTLC
   transactions (intuitively I think all of them, but at least HTLC-fulfilling transactions definitely).
 * Our poor B, being exploited, above, will attempt to spend their anchor output with a CPFP even if they aren't sure C
   has broadcast the HTLC-Success transaction! This is fine as B already knows the txid, and just wants to learn whats
   in the witness (assuming there is one).

For those from bitcoin-dev still reading who are thinking "blah, you clearly don't need anything else, you have a
solution!" we're talking about extra outputs out the wazoo for hopefully-unnecessary edge cases involving transactions
entering the mempool which a user wants to avoid confirming! This severely cuts into the lowest-value HTLCs which can be
sent "safely" and adds a significant social cost of extra low-value, possibly-uneconomical outputs in the chain.

Still, lacking a better idea, and with a strong desire to make lightning's security more practical in a world where
Bitcoin miners are paid to operate, we should probably start considering moving forward with this.

Thanks for reading,
Matt

[1] https://github.com/lightningnetwork/lightning-rfc/pull/688

[2] I'll note that while it is a "severe issue", given the general issues with fee-prediction described in background 3,
its pretty hard to argue its really in the scope of the security model of lightning today. If there were an easy fix to
it, we'd have deployed it by now in response to private discussion, but, sadly, there is not.