Skip to content
Crypto.world

Research

Your Loss is My Gain: Low Stake Attacks on Liquid Staking Pools

PoS blockchains are assumed to be economically secure because attacking consensus safety or liveness requires a large amount of stake (e.g., over 33%). The authors argue that liquid staking opens a…

Your Loss is My Gain: Low Stake Attacks on Liquid Staking Pools
arXiv
2605.01025
Published
2026-05-01
Authors
Sen Yang, Aviv Yaish, Arthur Gervais, Fan Zhang

AI summary

Overview

Research area: Blockchain economics and protocol security — specifically Proof-of-Stake (PoS) consensus-layer incentive attacks combined with application-layer (DeFi) market manipulation, in the context of liquid staking on Ethereum.

Technical level: Intermediate. The paper combines an empirical finance-style correlation study, a deep reinforcement learning (DRL) framework, and Monte Carlo simulation; readers need some familiarity with PoS consensus (leader election, fork choice) and basic statistics, but the core argument is accessible.

Scope (one sentence): The paper shows that an adversary holding a small share of stake can degrade a liquid staking pool's consensus-layer performance and profit at the application layer by shorting that pool's liquid staking token (LST), exploiting an empirically measured positive association between pool APR and subsequent LST returns.

What This Paper Is About

PoS blockchains are assumed to be economically secure because attacking consensus safety or liveness requires a large amount of stake (e.g., over 33%). The authors argue that liquid staking opens a different attack surface that these arguments do not capture: an adversary does not need to break safety or liveness at all — it only needs to make a target pool's validators perform worse, so that the pool's token loses value relative to ETH.

The paper's goal is to (1) empirically establish that staking pool performance is associated with subsequent LST returns, (2) construct and automatically discover consensus-layer strategies that degrade a chosen victim pool, and (3) show that application-layer positions — chiefly leveraged shorting — can monetize the resulting price movement profitably.

Key Contributions

  1. First empirical study linking liquid staking pool performance to LST prices. Using Ethereum data across six pools (Lido, Binance, Coinbase, Ether.fi, Rocket Pool, Mantle), the authors report consistently positive Pearson and Spearman correlations between a pool's APR and its 180-day forward, ETH-normalized LST return — moderate for most pools and strong for Coinbase and Binance.

  2. A cross-layer attack framework that separates the attack surface from the profit channel. Rather than violating safety or liveness, the adversary manipulates the consensus layer (leader election) to hurt a target pool's realized block allocation, and extracts profit at the application layer where LSTs are traded.

  3. A deep reinforcement learning framework for discovering leader-election manipulation strategies. Unlike prior work that maximizes the adversary's own block share via dynamic programming on an explicit MDP, this framework supports a victim-targeted objective and handles the state-space explosion. It recovers known self-optimization strategies as a validation baseline and finds new degradation-oriented behaviors, including an MEV-aware (heterogeneous block value) extension.

  4. A taxonomy of application-layer monetization channels plus a quantitative analysis of leveraged shorting. The authors classify four monetization routes (shorting, prediction markets, liquidation capture, cross-exchange arbitrage) and use Monte Carlo simulations calibrated to real-world parameters to estimate profitability of leveraged shorting.

Main Findings

  • Liquid staking dominates Ethereum stake: in the paper's stake distribution (as of Feb 2026, based on data from a cited validator-labeling dataset), liquid staking pools hold 49.0%, other stakers 27.9%, and centralized exchanges 23.1%; pooled and liquid staking together have comprised about 50% of stake since the Merge.

  • Pool performance is positively associated with future LST returns: all six pools show positive Pearson and Spearman correlations. Coinbase is strongest (Pearson 0.645, Spearman 0.708), Binance is relatively strong (0.434 and 0.516), Lido is moderate (0.391 and 0.387), Ether.fi is 0.407 and 0.302, Rocket Pool 0.176 and 0.260, and Mantle 0.112 and 0.315. All regression coefficients on contemporaneous APR are positive and statistically significant (p-values from <0.0001 to 0.0365), with the largest coefficients for Ether.fi (1.7247) and Coinbase (0.8719).

  • A low-stake adversary can disproportionately harm a victim pool: under the victim-targeted objective, an adversary with 20% stake can reduce the realized block allocation of another 20% stake pool by about 8.3%. The learned policies trade a limited adversarial loss for a disproportionate victim loss.

  • Degradation-oriented policies also hurt chain quality: compared with self-optimization policies, they more often sacrifice or displace blocks in order to damage the victim pool.

  • MEV changes learned behavior: with heterogeneous block values, self-optimization policies become more conservative about sacrificing the adversary's own high-value slots, while both self-optimization and victim-targeted policies become more willing to fork-displace victim blocks that carry valuable MEV opportunities.

  • Leveraged shorting can be profitable more often than not: Monte Carlo simulations calibrated to real-world data and parameters yield a likelihood of profit above one-half for three representative pools — Coinbase, Ether.fi, and Rocket Pool. Compared with roughly 3% APR from honest staking, expected profit from successful executions could be 11 times higher, providing a strong incentive even if attacks are executed infrequently.

  • A gap between consensus and economic security: the attacks remain viable even when the underlying consensus mechanism is provably economically secure, because the profit does not come from breaking consensus guarantees.

Methodology in Plain English

Step 1 — Measure the market link. The authors collect Ethereum data from block 15,537,393 to block 22,677,606 (the Merge on September 15, 2022 through June 10, 2025, i.e., 1,000 days of PoS). They gather daily USD prices of ETH and six LSTs from CoinGecko, map validators to pools using a dataset of 1,969,253 validators (1,887,716 of which carry pool tags) plus an address-to-pool-tag mapping covering 65 validator addresses across 43 pools, and compute each pool's APR from consensus-layer rewards (attestation, proposer, sync committee) and execution-layer rewards (MEV-Boost relay data for MEV-Boost blocks, on-chain block value otherwise), divided by daily effective balance and annualized. To strip out the common market factor, they normalize each LST price by the ETH price. They then correlate pool APR with 180-day forward log returns of the normalized price, reporting Pearson and Spearman correlations and regressions with heteroskedasticity- and autocorrelation-consistent (HAC) standard errors. Robustness checks at 90-day and 60-day horizons are reported in the appendix.

Step 2 — Learn the manipulation. They build a simulated Ethereum environment implementing leader election (RANDAO-based proposer selection, 32 slots per epoch) and the LMD-GHOST fork choice rule with a 40% proposer boost. An RL agent plays only the adversary's validators; all other validators follow the protocol. The agent's actions — propose, miss, hide, fork, regret — are compressed to three choices {0, 1, 2} because the only genuinely meaningful decision point is whether to fork or abandon a private chain once an honest competing block appears. To overcome sparse, delayed rewards (effects show up in the next epoch's block allocation), the observation vector is 7-dimensional and includes three counterfactual "oracles" (Ω_prop, Ω_skip, Ω_fork) that estimate the next-epoch block allocation for the adversary and the victim pool under each hypothetical action. A composite reward with weights β, ω, and γ lets them switch between self-optimization (ω = 0), pure griefing (β = 0), and intermediate objectives. An MEV-aware extension adds a bounded pool of high-value opportunities with arrival probability δ and added value F, plus value-aware state variables and reward terms.

Step 3 — Price the profit. They enumerate application-layer monetization channels and analyze leveraged shorting via Monte Carlo simulation: the adversary deposits neutral collateral, borrows and sells the target LST, and repurchases it later to repay the debt, subject to frictions such as liquidation risk, slippage, and borrowing cost. (The truncated text cuts off mid-description of the recursive leverage parameter, so the full simulation parameterization is not visible in the provided content.)

Why This Matters

Impact on research. The paper reframes PoS economic security: the standard argument bounds the cost of violating safety or liveness, but says nothing about adversaries who profit by degrading a subset of validators and trading against the resulting token price. It also introduces a methodological shift from explicit-MDP dynamic programming (used in prior leader-election biasing work) to deep RL, which handles objectives other than the adversary's own block share.

Real-world applications:

  • Staking pool and LST risk assessment: pools and their DAOs can use the measured performance-to-price sensitivity to judge how much operational degradation actually threatens their token's value.
  • DeFi risk management: lending protocols listing LSTs as collateral can account for the possibility of targeted, temporary performance degradation and the resulting liquidation cascades.
  • Market surveillance: exchanges and analytics firms can monitor for the specific consensus-layer signatures (block withholding, ex-ante reorgs displacing a single pool's blocks) that precede repricing.
  • Validator/operator monitoring: delegators can use realized block allocation relative to stake-proportional expectation as an operational health metric.

Industry relevance. With liquid staking pools holding 49.0% of Ethereum stake in the paper's snapshot, an attack surface affecting the largest staking category has direct implications for LST issuers, liquid staking operators, lending markets that accept LSTs as collateral, and prediction and derivatives venues. The claim that expected profit can be 11 times the roughly 3% honest-staking APR — with profit likelihood above one-half for Coinbase, Ether.fi, and Rocket Pool — is an incentive argument aimed squarely at practitioners, not just theorists.

Future Directions

  1. Detectability and mitigation. The paper documents the vulnerability but does not report a deployed defense. Open questions include whether the learned withholding and forking patterns are distinguishable from benign missed slots, and what protocol changes (e.g., to leader election randomness or proposer boost) would blunt victim-targeted degradation.

  2. Extending beyond leveraged shorting. The authors present a four-class taxonomy (shorting, prediction markets, liquidation capture, arbitrage) but quantify only leveraged shorting in depth; calibrating the other three, which involve MEV competition, is left open.

  3. Combining consensus-layer manipulation with other attack vectors. The paper's Table 2 lists insider attacks and network attacks as lower-stake alternatives and defers details to an appendix; measuring how these compose with the leader-election strategies — or with larger-stake reorganization attacks — remains unexplored in the main text.

  4. Generalization beyond Ethereum and beyond the six studied LSTs. The empirical analysis covers six pools over 1,000 days; whether the performance–price association and the profitability of attacks hold for smaller LSTs, other PoS chains, and rebasing LST designs is not reported.

Target Audience

Protocol researchers and security analysts working on PoS consensus and staking economics; liquid staking providers and their risk teams; DeFi risk managers at lending protocols and exchanges that list LSTs; and graduate students or quantitatively inclined practitioners interested in the intersection of mechanism design, reinforcement learning, and crypto markets. Readers looking for a purely theoretical treatment may find this empirically grounded and simulation-driven; beginners will need to consult the background section for leader election and fork choice concepts.

Authors’ abstract

Permissionless Proof-of-Stake (PoS) economic security is predicated on the high cost of violating consensus safety or liveness. We show that liquid staking introduces additional risks that are not captured by standard PoS economic security arguments. Through an empirical study of Ethereum data, we find that the operational performance of liquid staking pools is positively associated with subsequent normalized liquid staking token (LST) returns. Motivated by this, we present…

The full abstract on arXiv
Read the original paper