EIP-7942 Available Attestation
No call decision · Consensus & staking · EIP status Draft · created 2025-04-30
A proposal that recognizes a block as stable only if it includes attestations from at least one-third of the validators of the previous slot, and changes the fork choice rule to the longest stable chain rule to prevent chain reorg attacks.
Why it is needed
Every known effective attack on Ethereum proof-of-stake (PoS) is a reorg attack, that is, an attack that pushes an already-created block off the chain. The original text categorizes these into two branches: attacks that alter block weight and attacks that cause the block tree to be filtered. The response so far has been piecemeal, fixing one problem at a time as it surfaced. For example, proposer boosting, introduced to prevent ex-ante reorg attacks and balancing attacks, gave rise to a new attack called the sandwich reorg attack, in which two malicious proposers collude, and the response to that in the Capella upgrade led to the staircase attack. Many of the existing mitigations either lack formal proof or rely on the additional assumption that the attacker's stake is 20% or less. This proposal aims for a fundamental solution that is provably safe while remaining lightweight to implement.
What changes
Today
- Because the fork choice rule HLMD-GHOST selects the canonical chain based on block weight computed from attestation counts, it cannot prevent an attacker from manipulating things by withholding and then releasing attestations
- Responses to reorg attacks have accumulated as individual patches, such as proposer boosting and safe slots, and some of these have induced new attacks
- There is no stability criterion to confirm that a block has been sufficiently observed
After EIP-7942
- A block in slot t becomes a stable block when its parent is a block (AA) in slot t-1 that received attestations from more than one-third of validators
- The canonical chain is determined as the chain with the most stable blocks, and ties are broken by choosing the chain whose leaf block has the higher slot number
- Transactions in unstable blocks are preserved via an unstable_root reference so they can later be finalized together
- All known reorg attacks are provably blocked in a synchronous network
Who is affected
- Client teams: The parent selection rule in the validator specification and the fork choice specification must be modified, and because this is incompatible with existing clients, a hard fork or protocol version flag is required
- Validators: When proposing a block, first compute the AA block of the previous slot and use it as the parent; only when there is none is the fork choice result used as the parent
- Users and app developers: As reorg attacks become harder, the risk that a once-included transaction gets pushed off the chain decreases
- Protocol researcher: A companion paper accepted at USENIX Security 2025 contains the formal proof and an evaluation covering 16,384 validators
In depth
Weak quorum and the definition of a stable block
This borrows the concept of a weak quorum from BFT protocols. Among N validators, if at most f can be malicious, and f+1 have voted for block b, then at least one honest validator must have validated b, so these f+1 attestations serve as proof of availability for b. A block in slot t that receives attestations from more than one-third is denoted AA_t. A block b proposed in slot t is a stable block if its parent is AA_{t-1}, and otherwise an unstable block; a chain whose leaf block is a stable block is a stable chain.
Committee splitting and threshold ϑ calculation
In practice, validators are randomly split into 32 committees, so the number of malicious validators per committee is approximated by a binomial distribution. The threshold ϑ for the number of malicious validators within a committee is computed as ϑ = ⌊μ + σ·Φ⁻¹(1-p)⌋ for an allowed failure probability p, where μ = f' and σ² = f'(1 - f/n). Ethereum currently has about 2²⁰ validators, and at p = 10⁻⁶ the ratio is 0.3457, meaning the protocol fails once every 10⁶ slots (138 days). At p = 10⁻⁹, a malicious reorg occurs once every 380 years. As n grows large the ratio converges to f/n, so the conclusion is that ϑ can simply be set to one-third of the committee size.
Change 1: Parent selection rule
A block b at slot t must include ϑ+1 attestations proving that the parent is a valid output of slot t-1, in which case the parent is AA_{t-1}. If there is no AA_{t-1}, the output of the fork choice rule is used as the parent, and that block becomes an unstable block. In the validator specification, at proposal time the AA block is computed at the start of the slot, setting parent_root to the root aa_root of the later AA block; if there is no AA block, get_head(store), after running on_tick and applying pending attestations from slot-1, is used as parent_root.
Change 2: Replacing HLMD-GHOST with the longest stable chain rule
The fork choice is changed to the Longest Stable Chain rule, which outputs the head of the chain with the most stable blocks. Ties are broken by choosing the chain whose leaf block has the highest slot number. The implementation adds a Chain class with fields leaf (Root), length (uint64), and justified_checkpoint (Checkpoint), adds chains: Set[Chain] to Store, and get_head returns the leaf of the chain with the maximum length among the filtered chains. The reason for the replacement is that HLMD-GHOST is based on block weight determined by attestation counts, which cannot prevent an attacker from withholding attestations.
Change 3: Preserving transaction continuity via unstable_root
Adds an unstable_root field to the block. The proposer sets unstable_root to the hash of the latest unstable block, or leaves it as None if no such block exists. Transactions in unstable blocks must not conflict with transactions in the stable chain (the longest chain), and once a block containing unstable_root is finalized, the transactions within it can be finalized together. This is a mechanism to prevent transactions carried in unstable blocks from being discarded.
Change 4: Simplifying the filtering rule and justification
get_filtered_block_tree is simplified so that only chains justifying a checkpoint equal to the store's justified_checkpoint remain. The unrealized_justified_checkpoint and unrealized_finalized_checkpoint fields are removed, eliminating compute_pulled_up_tip, i.e. the handling of justification and finalization that occurs mid-epoch. A chain's justified_checkpoint is now updated only when crossing an epoch boundary.
Basis for reorg resistance and safety/liveness
The attacker has two possible strategies: immediately proposing a block that conflicts with the canonical chain, or building a block that extends the canonical chain and delaying its release. The former is blocked because a block extending a non-leaf block cannot receive attestations from honest validators and so becomes an unstable block. The latter is blocked because if two or more stable blocks are withheld, at least one of them has already been observed by all honest validators (Lemma 2 in the paper). Since the finality mechanism, Casper, is not modified, safety is preserved, and liveness is achieved because reorg resistance holds during synchronous periods, so that after GST some block eventually gets finalized. The proposal states that it introduces no new message types and therefore no new attack surface.
Deployment requirements and current status
This proposal modifies the fork choice logic and attestation format, so it may be incompatible with existing clients, and deployment requires a hard fork or protocol version flag. The source status is Draft, so the details of the specification may still change. No decision on fork inclusion has come out of core developer calls yet.