Personally, I would have wait a bit before to go public on this, like letting some implementations
increasing their CLTV deltas, but anyway, it's here now.
 
Mempool-pinning attacks were already discussed on this list [0], but what we found is you
can _reverse_ the scenario, where it's not the malicious party delaying confirmation of honest
party transactions but malicious deliberately stucking its own transactions in the mempool to avoid
confirmation of timeout. And therefore gaming inter-link timelock to provoke an unbalanced
settlement for the victim ("aka you pay forward, but don't get pay backward").

How much attacks are practical is based on how you can leverage mempool rules to pin your own
transaction. What you're looking for is a  _mempool-obstruction_ trick, i.e a way to get honest party
transaction being bounce off due to your transaction being already there.

Beyond disabling RBF on your transaction (with current protocol, not anchor proposal), there is
two likely candidates:
* BIP 125 rule 3: "The replacement transaction pays an absolute fee of at least the sum paid by the original transactions."
* BIP 125 rule 5: "The number of original transactions to be replaced and their descendant transactions which will be evicted from the mempool must not exceed a total of 100 transactions."

Let's go through whole scenario:
* Mallory and Eve are colluding
* Eve and Mallory are opening channels with Alice, Mallory do a bit of rebalancing
to get full incoming capacity, like receiving funds on an onchain address through another Alice
link
* Eve send a HTLC #1 to Mallory through Alice expirying at block 100
* Eve send a second HTLC #2 to Mallory through Alice, expirying at block 110 on outgoing link
(A<->M), 120 on incoming link (E<->A)
* Before block 100, without cancellation from Mallory, Alice will force-close channel and broadcast
her local commitment and HTLC-timeout to get back HTLC #1
* Alice can't broadcast HTLC-timeout for HTLC #2 as it's only expires at 110
* Mallory can broadcast its Pinning Preimage Tx on offered HTLC #2 output on Alice's transaction,
feerate is maliciously chosen to get in network mempools but never to confirm. Absolute fee must
be higher than HTLC-timeout #2, a fact known to Mallory. There is no p2p race.
* As Alice doesn't watch the mempool, she is never going to learn the preimage to redeeem incoming
HTLC #2
* At block 110, Alice is going to broadcast HTLC-timeout #2, feerate may be higher but as absolute
fee is lower, it's going to be rejected from network mempools as replacement for Pinning Preimage
Tx (BIP 125 rule 3)
* At block 120, Eve closes channel and HTLC-timeout HTLC #2
* Mallory can RBF its Pinning Preimage Tx by a high-feerate one and get it confirmed

New anchor_output proposal, by disabling RBF, forces attacker to bid on the absolute fee. It may
be now a risk to loose the fee if Pinning Tx is confirming. You may extend your "pinning
lease" by ejecting your malicious tx, like conflicting or trimming out of the mempool one of its
parents. And then reannounce your preimage tx with a lower-feerate-but-still-high-fee before a
new block and a honest HTLC-timeout rebroadcast.

AFAICT, even with anchor_output deployed, even assuming empty mempools, success rate and economic
rationality of attacks is finding such cheap, reliable "pinning lease extension" trick.

I think any mempool watching mitigation is at best a cat-and-mouse hack. Contrary to node
advancing towards a global blockchain view thanks to PoW, network mempools don't have a convergence
guarantee. This means,  in a distributed system like bitcoin, node don't see events in the same
order, Alice may observe tx X, tx Y, tx Z and Bob may observe tx Z, tx X, tx Y. And order of events
affects if a future event is going to be rejected or not, like if tx Z disable-RBF and tx X try to
replace Z, Alice accepts X and Bob rejects it. And this divergence may perserve until a new block.

Practically, it means an attacker can provoke a local conflict to bounce off HTLC preimage tx out
of your mempool while broadcasting preimage tx without conflict to the rest of the network by
tweaking tx-relay protocol and so easily manipulating order of events for every node. A local
conflict is easy to provoke, just make tx A double-spent by both HTLC-preimage-tx and non-RBF-tx-B.
Announce txA+txB to mempool victim and txA+HTLC-preimage-tx to rest of network. When rest of
network announce HTLC-preimage-tx, it's going to rejected by your mempool.

Provoking local conflict assumes of course _interlayer_ mapping by an attacker, i.e mapping your LN
node to your full-node(s). Last time, we check, there was 982 match by IP for 4,500 LN/52,000
full-node. Mapping heuristics is an ongoing research subject and sadly seems affordable.

Yes a) you can enable full-RBF on your local node but blinding conflicting may still be with higher
feerate as everything is attacker malleable b) you may want to catch tx and extract preimage
on the p2p wire, but processing raw transaction would be such a DoS vector...

Overall, I think we all agree on the long term direction to get a Contracting-Protocols-Enhanced
mempool with a multiparty-safe-API, bundled with package relay deployment. Even if there is current
move toward this direction, this may take longer than expected as with any critical-safety
component in Core.

A temporary fix could be to resuscitate old work to ensure peering through a full-RBF propagation path,
but p2p implications are hard to gauge, like wouldn't guarantee p2p censorship resistance of this...

It's quite a tangled issue, with a good deal of both bitcoin and lightning knowledge so feel free
to verify and double-check more than usual

Cheers

[0] https://lists.linuxfoundation.org/pipermail/lightning-dev/2019-October/002240.html

Le mer. 22 avr. 2020 à 02:08, ZmnSCPxj via bitcoin-dev <bitcoin-dev@lists.linuxfoundation.org> a écrit :
Good morning Laolu, Matt, and list,


> >  * With `SIGHASH_NOINPUT` we can make the C-side signature
> >  `SIGHASH_NOINPUT|SIGHASH_SINGLE` and allow B to re-sign the B-side
> >  signature for a higher-fee version of HTLC-Timeout (assuming my cached
> >  understanding of `SIGHASH_NOINPUT` still holds).
>
> no_input isn't needed. With simply single+anyone can pay, then B can attach
> a new input+output pair to increase the fees on their HTLC redemption
> transaction. As you mention, they now enter into a race against this
> malicious ndoe to bump up their fees in order to win over the other party.

Right, right, that works as well.

>
> If the malicious node uses a non-RBF signalled transaction to sweep their
> HTLC, then we enter into another level of race, but this time on the mempool
> propagation level. However, if there exists a relay path to a miner running
> full RBF, then B's higher fee rate spend will win over.

Hmm.

So basically:

* B has no mempool, because it wants to reduce its costs and etc.
* C broadcasts a non-RBF claim tx with low fee before A->B locktime (L+1).
* B does not notice this tx because:
  1.  The tx is too low fee to be put in a block.
  2.  B has no mempool so it cannot see the tx being propagated over the P2P network.
* B tries to broadcast higher-fee HTLC-timeout, but fails because it cannot replace a non-RBF tx.
* After L+1, C contacts the miners off-band and offers fee payment by other means.

It seems to me that, if my cached understanding that `<0> OP_CHECKSEQUENCEVERIFY` is sufficient to require RBF-flagging, then adding that to the hashlock branch (2 witness bytes, 0.5 weight) would be a pretty low-weight mitigation against this attack.

So I think the combination below gives us good size:

* The HTLC-Timeout signature from C is flagged with `OP_SINGLE|OP_ANYONECANPAY`.
  * Normally, the HTLC-Timeout still deducts the fee from the value of the UTXO being spent.
  * However, if B notices that the L+1 timeout is approaching, it can fee-bump HTLC-Timeout with some onchain funds, recreating its own signature but reusing the (still valid) C signature.
* The hashlock branch in this case includes `<0> OP_CHECKSEQUENCEVERIFY`, preventing C from broadcasting a low-fee claim tx.

This has the advantages:

* B does not need a mempool still and can run in `blocksonly`.
* The normal path is still the same as current behavior, we "only" add a new path where if the L+1 timeout is approaching we fee-bump the HTLC-Timeout.
* Costs are pretty low:
  * No need for extra RBF carve-out txo.
  * Just two additional witness bytes in the hashlock branch.
* No mempool rule changes needed, can be done with the P2P network of today.
  * Probably still resilient even with future changes in mempool rules, as long as typical RBF behaviors still remain.

Is my understanding correct?

Regards,
ZmnSCPxj

>
> -- Laolu
>
> On Tue, Apr 21, 2020 at 9:13 PM ZmnSCPxj via bitcoin-dev <bitcoin-dev@lists.linuxfoundation.org> wrote:
>
> > Good morning Matt, and list,
> >
> > >     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.
> >
> > My (cached) understanding is that, since RBF is signalled using `nSequence`, any `OP_CHECKSEQUENCEVERIFY` also automatically imposes the requirement "must be RBF-enabled", including `<0> OP_CHECKSEQUENCEVERIFY`.
> > Adding that clause (2 bytes in witness if my math is correct) to the hashlock branch may be sufficient to prevent C from making an RBF-disabled transaction.
> >
> > But then you mention out-of-band agreements with miners, which basically means the transaction might not be in the mempool at all, in which case the vulnerability is not really about RBF or relay, but sheer economics.
> >
> > The payment is A->B->C, and the HTLC A->B must have a larger timeout (L + 1) than the HTLC B->C (L), in abstract non-block units.
> > The vulnerability you are describing means that the current time must now be L + 1 or greater ("A could claim the HTLC from B via the timeout mechanism", meaning the A->B HTLC has timed out already).
> >
> > If so, then the B->C transaction has already timed out in the past and can be claimed in two ways, either via B timeout branch or C hashlock branch.
> > This sets up a game where B and C bid to miners to get their version of reality committed onchain.
> > (We can neglect out-of-band agreements here; miners have the incentive to publicly leak such agreements so that other potential bidders can offer even higher fees for their versions of that transaction.)
> >
> > Before L+1, C has no incentive to bid, since placing any bid at all will leak the preimage, which B can then turn around and use to spend from A, and A and C cannot steal from B.
> >
> > Thus, B should ensure that *before* L+1, the HTLC-Timeout has been committed onchain, which outright prevents this bidding war from even starting.
> >
> > The issue then is that B is using a pre-signed HTLC-timeout, which is needed since it is its commitment tx that was broadcast.
> > This prevents B from RBF-ing the HTLC-Timeout transaction.
> >
> > So what is needed is to allow B to add fees to HTLC-Timeout:
> >
> > * We can add an RBF carve-out output to HTLC-Timeout, at the cost of more blockspace.
> > * With `SIGHASH_NOINPUT` we can make the C-side signature `SIGHASH_NOINPUT|SIGHASH_SINGLE` and allow B to re-sign the B-side signature for a higher-fee version of HTLC-Timeout (assuming my cached understanding of `SIGHASH_NOINPUT` still holds).
> >
> > With this, B can exponentially increase the fee as L+1 approaches.
> > If B can get HTLC-Timeout confirmed before L+1, then C cannot steal the HTLC value at all, since the UTXO it could steal from has already been spent.
> >
> > In particular, it does not seem to me that it is necessary to change the hashlock-branch transaction of C at all, since this mechanism is enough to sidestep the issue (as I understand it).
> > But it does point to a need to make HTLC-Timeout (and possibly symmetrically, HTLC-Success) also fee-bumpable.
> >
> > Note as well that this does not require a mempool: B can run in `blocksonly` mode and as each block comes in from L to L+1, if HTLC-Timeout is not confirmed, feebump HTLC-Timeout.
> > In particular, HTLC-Timeout comes into play only if B broadcast its own commitment transaction, and B *should* be aware that it did so --- there is still no need for mempool monitoring here.
> >
> > Now, of course this only delays the war.
> > Let us now consider what C can do to ensure that the bidding war will happen eventually.
> >
> > * C can bribe a miner to prevent HTLC-Timeout from confirming between L and L+1.
> >   * Or in other words, this is a censorship attack.
> >     * The Bitcoin censorship-resistance model is that censored transactions can be fee-bumped, which attracts non-censoring miners to try their luck at mining and evict the censoring miner.
> >       * Thus, letting B bump the fee on HTLC-Timeout is precisely the mechanism we need.
> >       * This sets up a bidding war between C requesting miners to censor, vs. B requesting miners to confirm, but that only sets the stage for a second bidding war later between C and B, thus C is at a disadvantage: it has to bribe miners to censor continuously from L to L+1 *and* additional bribe miners to confirm its transaction after L+1, whereas B can offer its bribe as being something that miners can claim now without waiting after L+1.
> >
> > The issue of course is the additional output that bloats the UTXO set and requires another transaction to claim later.
> > And if we have `SIGHASH_NOINPUT`, it seems to me that Decker-Russell-Osuntokun sidesteps this issue as well, as any timed-out HTLC can be claimed with a fee-bumpable transaction directly without RBF-carve-out.
> > (As well, it seems to me that, if both nodes support doing so, a Poon-Dryja channel can be upgraded, without onchain activity, to a Decker-Russell-Osuntokun channel: sign a transaction spending the funding tx to a txo that has been set up as Decker-Russell-Osuntokun, do not broadcast that transaction, then revoke the latest Poon-Dryja commitment transactions, then switch the mechanism over to Decker-Russell-Osuntokun; you still need to monitor for previous Poon-Dryja commitment transactions, but HTLCs now sidestep the issue under discussion here.)
> >
> > Regards,
> > ZmnSCPxj
> > _______________________________________________
> > bitcoin-dev mailing list
> > bitcoin-dev@lists.linuxfoundation.org
> > https://lists.linuxfoundation.org/mailman/listinfo/bitcoin-dev


_______________________________________________
bitcoin-dev mailing list
bitcoin-dev@lists.linuxfoundation.org
https://lists.linuxfoundation.org/mailman/listinfo/bitcoin-dev