Skip to content
Crypto.world

Research

Sandwiched and Silent: Behavioral Adaptation and Private Channel Exploitation in Ethereum MEV

Sandwich attacks — where an attacker places trades before and after a victim's swap to profit from the induced price movement — are usually discussed as a public-mempool problem.

Sandwiched and Silent: Behavioral Adaptation and Private Channel Exploitation in Ethereum MEV
arXiv
2512.17602
Published
2025-12-19
Authors
Davide Mancino, Davide Rezzoli

AI summary

Overview

Research area: Blockchain security and mechanism design — specifically Maximal Extractable Value (MEV) on Ethereum under Proposer–Builder Separation (PBS), with a focus on user behavior after sandwich attacks and on exploitation that occurs inside private order-flow channels.

Technical level: Intermediate. Readers should be comfortable with Ethereum transaction mechanics, mempools, DEX routers, and basic survival-analysis vocabulary, but the paper is largely empirical and readable without deep protocol expertise.

Scope (1 sentence): Using transaction-level Ethereum data from November 2024 to February 2025 enriched with mempool-visibility labels and ZeroMEV MEV classifications, the paper measures how sandwich victims react and switch to private routing, and detects and quantifies sandwich attacks that occur on private transaction paths.

What This Paper Is About

Sandwich attacks — where an attacker places trades before and after a victim's swap to profit from the induced price movement — are usually discussed as a public-mempool problem. The paper asks two linked questions: what do users actually do after they get sandwiched (do they come back, do they leave, do they switch to private routing?), and does private routing actually protect them?

The authors find that most victims keep transacting, a substantial minority migrate to private channels, and private channels themselves host thousands of confirmed sandwich attacks concentrated in the hands of very few operators. The headline implication is that private order flow is a narrower but more concentrated attack surface, not a shield.

Key Contributions

  1. A windowed, n-indexed survival analysis of post-attack behavior. The authors invert the usual Kaplan-Meier framing to measure reactivation (resuming on-chain activity within 60 days of the n-th public sandwich) and first-time adoption of private routing, defining non-reactivating users as "churned" and users with n > 1 as experiencing "repeated exposure."

  2. First systematic detection and economic quantification of private-path sandwiches. They report 2,932 confirmed private sandwich attacks affecting 3,126 private victim transactions in November–December 2024, with $409,236.97 in victim-side losses and $293,785.95 in attacker profits.

  3. Evidence that private routing does not prevent MEV extraction. The private-path results directly challenge the assumption that routing away from the public mempool offers genuine protection.

  4. Documentation of extreme attacker concentration. One bot (address 0xae2f…ae13) executes 1,901 private frontruns, roughly 65% of private frontrun volume, which the authors argue is consistent with potential compromise of, or collusion within, private routing infrastructure.

Main Findings

  • Migration to private routing is substantial. Of 36,205 users in the dataset, 13,473 (37.2%) transitioned to private channels within 60 days of their first public sandwich; the abstract rounds this to "around 40%." Among the 34,102 users who reactivated in the window, adoption reaches 39.5%.

  • Adoption rises with repeated exposure (a learning effect). Adoption climbs from approximately 37% after the first attack to over 54% by the seventh attack. The trend holds among reactivated users as well, at consistently higher levels.

  • Churn is front-loaded and small. Non-reactivation peaks at about 7.5% after the first sandwich and falls to 1–2% for repeatedly attacked users — a pattern the authors attribute to survivor bias. Users who endured up to ten sandwiches show reactivation rates exceeding 98%.

  • Most reactivation is immediate. A large share of addresses resume activity the same day as the attack, with the steepest cumulative reactivation occurring within the first 24–48 hours, suggesting users are either unaware of MEV or treat it as routine overhead.

  • Loss size does not explain who switches. Loss distributions for switchers and non-switchers are broadly similar; a Mann-Whitney U test detects a difference only because of the large sample, with Cliff's δ ≈ 0.036 (practically indistinguishable). The authors conclude adoption is driven by unobserved individual factors rather than systematically higher damage.

  • Private sandwiches are real and economically material. In November–December 2024: 2,932 private attacks, 3,126 private victim transactions, $409,236.97 in victim losses and $293,785.95 in attacker profits. Losses are heavy-tailed: mean $137.47, median $26.85, interquartile range [$12.09, $69.16], maximum $2,744.69. Profits: mean $106.95, median $16.35, IQR [$7.47, $43.05], maximum $37,965.21.

  • Private exploitation is highly concentrated among attackers. The top bot alone handles 1,901 private frontruns (~65%); the second-largest attacker handles only 229, followed by 82, 67, 53, 41, 29, 28, 25, and 24. The paper notes Etherscan associates the leading address with the ENS name jaredfromsubway.eth, reported for context only.

  • Victims are repeatedly targeted. The most-attacked private victims experienced 28, 23, and 22 separate private sandwich events, indicating sustained predatory behavior against a subset of users.

  • Targeting is skewed toward a few routers and aggregators. Destination 0x3fc9…7fad concentrates 856 victim transactions (27% of the total), followed by aggregator 0x1111…0582 with 467, Uniswap V2 router 0x7a25…488d with 397, and 0x0000…2734 with 356. The trading interface Banana Gun (0x3328…9c49) accounts for 48 victim transactions, showing the pattern extends to emerging trading tools.

  • Overall public/private sandwich counts. After filtering: 123,030 public victim events and 3,126 private victim events; 117,581 confirmed public sandwich attacks and 2,932 confirmed private attacks.

  • Private routing share nearly doubled over the period. Private transactions were ~31.8% in November 2024, ~35.5% in December, 34.8% in January 2025 (13,189,288 / 37,865,791), and 50.1% in February (17,807,840 / 35,574,446).

  • Fork-induced visibility leakage is negligible. Only 8 private victim transactions fall in blocks later marked as forked — below 0.3% of the private-sandwich sample — and removing them changes no aggregate counts or distributional patterns.

Methodology in Plain English

The authors assemble a transaction-level panel for November 2024 through February 2025 from XBlock-ETH, covering explicit block ranges for each month. Each transaction hash is labeled public if it was observed in the mempool before inclusion (cross-referenced against MempoolDumpster) and private otherwise. Each hash is then enriched with ZeroMEV's mev_type label, protocol, user-loss and extractor-profit estimates, and swap volumes.

Sandwich detection works block by block. Transactions are sorted by block number and transaction index to fix execution order. For each transaction labeled a frontrun by ZeroMEV, sent by address s, the authors locate the first later transaction in the same block also sent by s and labeled a backrun. Transactions strictly between the two form a candidate victim region. A sandwich is confirmed only if at least one transaction in that region carries ZeroMEV's sandwich label (which denotes the victim). Self-sandwiching is excluded by requiring the victim's sender to differ from the attacker, and only distinct external victims are retained. If any confirmed victim is private, the frontrun sender is recorded as a private-path attacker.

For behavior, the unit of analysis is the address and the index n of its n-th sandwich. The authors track whether each address transacts again within 60 days (reactivation) and whether it makes a first transition to private routing in that window, plotting cumulative incidence curves in the style of survival analysis but inverted to measure recovery rather than cessation. Note that detailed ZeroMEV labels were collected only for November and December 2024, so the visibility-by-MEV breakdown and all private-path economics cover those two months.

Why This Matters

For research, the paper reframes private order flow from a mitigation to a measurement problem. It supplies a reproducible detection pipeline for private-path sandwiches, an empirical baseline (2,932 attacks, $409,236.97 in losses over two months) for future work, and behavioral evidence that migration decisions are not explained by loss magnitude — a puzzle for models of user sophistication under MEV.

Real-world applications:

  • Wallet and RPC providers can use the reactivation and adoption curves to time user-education prompts, since the first 24–48 hours and the first attack are when behavior is most malleable.
  • DEX routers and aggregators named as concentrated targets (0x3fc9…7fad with 856 victim transactions, 0x1111…0582 with 467, the Uniswap V2 router with 397, plus Banana Gun) have a direct incentive to add slippage or ordering defenses on their highest-traffic entry points.
  • Private relay and builder operators can treat the single-operator concentration (~65% of private frontruns) as an infrastructure-audit trigger.
  • Risk and compliance teams at trading firms can incorporate private-path exposure into execution-quality monitoring rather than assuming private means safe.

Industry relevance: the growth of private routing from ~31.8% to 50.1% of transactions means the MEV market's protective layer is becoming the dominant path for order flow. If that path is exploitable and concentrated, the systemic-risk profile of Ethereum execution changes — which matters to builders, relays, proposers, and to ongoing protocol work on enshrined PBS.

Future Directions

  1. Extend MEV-labeled coverage beyond November–December 2024. The detailed ZeroMEV enrichment covers only two of the four months, so whether private-path concentration and economics hold through January–February 2025 (when private share hit 50.1%) is unmeasured.

  2. Explain the adoption decision. Since loss distributions for switchers and non-switchers are practically indistinguishable (Cliff's δ ≈ 0.036), the drivers of migration remain unidentified; individual-level factors such as perceived risk or tooling exposure need direct study.

  3. Investigate how private transactions leak. The single bot with 1,901 private frontruns raises an open question about whether this reflects infrastructure compromise, privileged access, or collusion among intermediaries — a question the paper raises but does not resolve.

  4. Evaluate protocol-level defenses empirically. The authors call for commit-and-reveal schemes, encrypted intents, and careful evaluation of TEE-based architectures, with the implication that these need to be benchmarked against the private-path baseline established here.

Target Audience

Researchers and practitioners in blockchain security, MEV, and mechanism design; builder/relay/proposer infrastructure operators; wallet, RPC, and DEX-router engineers; and DeFi risk analysts. It is also useful for quantitatively inclined readers in market microstructure who want an empirical case study of how users respond to predatory execution and whether the private alternative actually works. Readers without familiarity with Ethereum transaction ordering will need background reading on mempools, PBS, and sandwich mechanics.

Authors’ abstract

How users adapt after being sandwiched remains unclear; this paper provides an empirical quantification. Using transaction level data from November 2024 to February 2025, enriched with mempool visibility and ZeroMEV labels, we track user outcomes after their n-th public sandwich: (i) reactivation, i.e., the resumption of on-chain activity within a 60-day window, and (ii) first-time adoption of private routing. We refer to users who do not reactivate within this window as…

The full abstract on arXiv
Read the original paper