Hi,

I'm writing to report a defect in Bitcoin Core bip125 logic with minor security and operational implications for downstream projects. Though this defect grieves Bitcoin Core nodes 0.12.0 and above, base layer safety isn't impacted.

# Problem

Bip 125 specification describes the following signalling mechanism :

"
This policy specifies two ways a transaction can signal that it is replaceable.

* Explicit signaling: A transaction is considered to have opted in to allowing replacement of itself if any of its inputs have an nSequence number less than (0xffffffff - 1).

* Inherited signaling: Transactions that don't explicitly signal replaceability are replaceable under this policy for as long as any one of their ancestors signals replaceability and remains unconfirmed.

One or more transactions currently in the mempool (original transactions) will be replaced by a new transaction (replacement transaction) that spends one or more of the same inputs if,

# The original transactions signal replaceability explicitly or through inheritance as described in the above Summary section.
"

An unconfirmed child transaction with nSequence = 0xff_ff_ff_ff spending an unconfirmed parent with nSequence <= 0xff_ff_ff_fd should be replaceable as the child transaction signals "through inheritance". However, the replacement code as implemented in Core's `PreChecks()` shows that this behavior isn't  enforced and Core's mempool rejects replacement attempts of an unconfirmed child transaction.

Branch asserting the behavior is here : https://github.com/ariard/bitcoin/commits/2021-03-test-rbf

# Solution

The defect has not been patched.

# Downstream Projects Affected

* LN : State-of-the-art pinning attacks against second-stage HTLCs transactions were thought to be only possible by exploiting RBF rule 3 on the necessity of a higher absolute fee [0]. However, this replacement defect opens the way for an attacker to only pin with an opt-out child without a higher fee than the honest competing transaction. This lowers the cost of attack as the malicious pinning transaction only has to be above mempools'min feerate. This also increases odds of attack success for a reduced feerate diminishes odds of confirmation ending the pinning.

A functional test demo illustrating cases is available on this branch: https://github.com/ariard/bitcoin/commits/2021-05-htlc-preimage-pinnings

LN nodes operators concerned by this defect might favor anchor outputs channels, fully mitigating this specific pinning vector.

* Onchain DLC/Coinswap/Vault : Those contract protocols have also multiple stages of execution with time-sensitive transactions opening the way to pinning attacks. Those protocols being non-deployed or in early phase, I would recommend that any in-protocol competing transactions explicitly signal RBF.

* Coinjoin/Cut-Through : if CPFP is employed as a fee-bumping strategy, if the coinjoin transaction is still laying in network mempools, if a fee-bumping output is spendable by any protocol participant, this fee-bumping mechanism might be halted by a malicious protocol participant broadcasting an low-feerate opt-out child. According to bip125, if the coinjoin parent tx signals replaceability, the child transaction should be replaceable, whatever its signaling. However Core doesn't apply this policy. RBF of the coinjoin transaction itself should be used as a fallback. I'm not aware of any deployed coinjoin using such "anyone-can-bump" fee-bumping strategy.

* Simple wallets : RBF engines' behaviors might be altered in ways not matching the intent of their developers. I invite RBF engines dev to verify what those components are doing in the light of disclosed information.

# Discovery

While reviewing the LN dual-funding flow, I inquired on potential new DoS vectors introduced by relying on counterparty utxos in this following analysis [1]. The second DoS issue "RBF opt-out by a Counterparty Double-Spend" is relying on a malicious chain of transactions where the parent is signaling RBF opt-in through nSequence<=0xff_ff_ff_ff-1 but the child, servicing as a pinning transaction, opt-out from the RBF policy. This pinning trick conception was matching my understanding of Core code but while reading again the specification, I observed that it was inconsistent from the inherited signaling mechanism as described in the bip's "Summary" section.

After exercising the logic, I did submit the defect to Dave Harding, asking confirmation of divergence between Bitcoin Core and BIP 125. Soon after, he did confirm it and pointed that the defect has been there since the 2015's PR introducing the opt-in RBF, advicing to to consider security implications for deployed second-layer protocols. After noticing the minor implications for pinning attacks on second-stage LN transactions while talking with Matt Corallo, I did disclose to the Bitcoin Core security list.

My initial report was recommending avoiding a covert patch in the mempool as risks of introducing DoS in this part of the codebase seemed to outweigh security of deployed LN channels. This direction was agreed by the opinions expressed on the security list. Beyond, there was a lack of agreement on how to proceed with the disclosure as so far in the history project, transaction relay policy have not been considered as strongly reliable. Though from now on, L2 protocols like Lightning are making assumptions on subset of this policy for their safety, such as the highlighted RBF one.

Defect was disclosed to the LN projects maintainers, informing them that currently in deployment anchor outputs protocol upgrade was mitigating against this defect though old channels will stay vulnerable. To the best of my knowledge, I didn't identify other deployed protocols of which coins safety are impacted by this defect.