Skip to content
Crypto.world

Research

No Place to Hide: An Analysis on Protected Order Flow Sandwich Attacks

Ethereum's public mempool became a "dark forest" where sandwich bots front-run and back-run ordinary trades. Chains have responded with private RPCs, order-flow auctions, local mempools…

No Place to Hide: An Analysis on Protected Order Flow Sandwich Attacks
arXiv
2609.28115
Published
2026-09-23
Authors
Lioba Heimbach, Ozan Solmaz, Burak Öz, Christof Ferreira Torres

AI summary

Overview

Research area: Blockchain security and MEV (maximal extractable value), specifically transaction ordering attacks on decentralized exchanges.

Technical level: Intermediate. The paper assumes some familiarity with mempools, block builders, proposers, and AMM pools, but its core argument is accessible.

Scope: A three-year (1 July 2023 to 30 June 2026) longitudinal measurement of sandwich attacks that succeed against transactions submitted through mechanisms intended to prevent front-running, across Ethereum, Solana, Tron, Base, Arbitrum, and Monad.

What This Paper Is About

Ethereum's public mempool became a "dark forest" where sandwich bots front-run and back-run ordinary trades. Chains have responded with private RPCs, order-flow auctions, local mempools, first-come-first-served ordering, and centralized sequencers, collectively what the authors call protected order flow. This paper asks a simple but largely unexamined question: do these protections actually work? The authors measure sandwich attacks against protected transactions across six chains, count how much attackers earn, and trace each failure back to the specific stage of the transaction pipeline where the protection breaks down.

Key Contributions

  1. The first longitudinal measurement study of sandwich attacks across six chains with heterogeneous front-running protections, spanning three years and covering six of the ten largest chains by TVL as of 17 August 2026.
  2. New detection heuristics that capture wide attacks (within-block and cross-block) rather than only tight sandwiches, plus a persistence filter that separates sandwich bots from high-frequency trading that happens to produce sandwich-shaped patterns.
  3. A three-year measurement identifying 28,042,725 protected order flow sandwiches on Solana, 38,567 on Tron, 30,607 on Ethereum, and 1,889 on Base, plus 2,576 additional sandwiches on Ethereum enabled by reorgs, together yielding attackers over $346M in net profit.
  4. An attribution of these attacks to concrete exposure points: order-flow auctions and reorged blocks on Ethereum, first-come-first-served ordering under a public mempool on Tron, validator- and application-level exposure on Solana, and an RPC leak plus predictable victim behavior on Base.

Main Findings

  • Attacks succeed everywhere except two chains. Persistent sandwich bots exist on Solana (8,631 bots), Tron (12), Ethereum (7), and Base (4). Arbitrum and Monad show zero persistent sandwich attackers despite protections that superficially resemble Base's and Solana's respectively. The authors stress that attack counts reflect what attackers do today, not what is theoretically possible.
  • Profit is overwhelmingly concentrated on Solana. Solana attackers extracted $383,433,932 gross and $345,185,014 net, with fees taking 10.0% of gross. By contrast Ethereum's 7 bots extracted $622,358 gross and $254,566 net (59.1% to fees), Tron's 12 bots $901,069 gross and $642,542 net (28.7% to fees), and Base's 4 bots $3,031 gross and $2,416 net (20.3% to fees).
  • Reorgs leak private transactions. 2,576 additional sandwiches by 93 bots landed on 2,875 private Ethereum transactions that a reorged block likely made public. For the 45.7% of these attacks that could be priced, gross profit was $279,848 and net $171,777.
  • Protected sandwiches are structurally different from public-mempool ones. Only 0.25% of Ethereum's protected sandwiches are tight, whereas public-mempool sandwiches on Ethereum are mostly tight, with a median of 1 transaction between legs and 99.4% within 6 transactions. The median gap from front-run to first victim and victim to back-run is 58 and 131 transactions on Ethereum, 13 and 23 on Tron, 2 and 1 on Base, and 1 and 1 on Solana, with 99th percentiles of 314 on Base and 8,651 on Solana.
  • Lower success rates reflect blind placement. Profitable attacks range from 88.0% on Solana to 97.2% on Base, versus 98.9% for Ethereum public-mempool sandwiches. Interference explains much of the gap: an unrelated swap in the attacked pool between the legs appears in 96.2% of unprofitable versus 22.4% of profitable attacks on Ethereum, 91.0% versus 6.5% on Tron, and 88.7% versus 2.0% on Base.
  • Ethereum's protected sandwiching is one entity. Seven bots account for all 30,607 attacks, and the two largest (13,661 and 10,202 sandwiches) cover 78.0% of them. The bots are active in mostly non-overlapping windows with week-to-week handovers, and bytecode similarity plus deployment history indicate all seven belong to a single entity.
  • Order-flow auctions are a plausible exposure path. Of 34,360 victims across 28,128 successful private attacks, 66.8% appear in at least one OFA dataset and 35.5% in multiple. 53.0% of attributed victims appear in multiple OFA datasets, and 28.3% appear in both MEV-Share and MEVBlocker — a combination the authors argue can be correlated, since a searcher can double-hash MEVBlocker hashes and match them against the MEV-Share stream. 88.6% of these victims are sole targets of their sandwich, making incidental inclusion unlikely.
  • Full-disclosure auctions carry a different risk. Blink appears in 62.2% of attributed victims and 33.4% are Blink-only, with 89.8% of Blink-only victims in single-victim sandwiches. Because Blink discloses full transaction data to whitelisted searchers, no cross-feed inference is needed to size a sandwich.
  • Many victims had no viable back-run alternative. Of 39,461 victim transactions, 68.1% had only one active direct pool and 57.8% had an endpoint proxy of one. For the 31,259 victims front-run in the same block, the authors rule out a profitable back-run for 19,410 (62.1%) under their coverage.
  • Solana's exposure shifted layers over time. In the first half of the dataset a subset of validators accounts for a disproportionate share of attacks; from 2025 onward this signal weakens and victim concentration shifts toward specific applications, making the application layer the dominant exposure source.
  • Base's attacks correspond to short-lived, distinct vulnerabilities. Activity is confined to three bursts: 16 days in August 2023 averaging 13 attacks per day (56.4% tight), eight days in February 2025 at 30 per day (entirely cross-block wide), and two weeks in April 2026 accounting for 76.7% of all Base attacks (76.0% within-block-wide, 23.8% tight). One bot coincides with a documented RPC-level incident that leaked pending transactions; others exploit highly predictable victim behavior.
  • Tron's FCFS ordering is defeated by latency. Tron retains a public mempool but orders by arrival time, which it claims prevents around 98% of front-running. The observed attacks are consistent with bots racing front-runs to the producer after seeing the victim, averaging 62 attacks per active day from late 2024 to mid-2026, with 82.9% within-block wide and 17.0% cross-block wide.

Methodology in Plain English

The authors collect on-chain data from the largest AMMs on each chain — Uniswap V2–V4 and forks such as SushiSwap, PancakeSwap, and SunSwap on EVM chains, Aerodrome V1 on Base, and Raydium, Orca Whirlpools, Meteora, and Pump.fun AMM on Solana — and release the full collection and detection pipeline as open-source code.

Detection uses three layered heuristics applied consistently across chains:

  • H1 (Structure): Find addresses that trade in both directions on a pool, aggregate consecutive same-direction swaps into a front leg and the following opposite-direction run into a back leg, and require the back leg to resell what the front leg bought within a 1% tolerance (ε = 0.01) and within five blocks on EVM chains or two leader rotations on Solana. At least one swap on the same pool in the front direction, by another wallet, must sit strictly between the legs; each such swap counts as a victim.
  • H2 (One Role): Prevent a transaction from playing multiple roles, and disqualify victims that also serve as candidate legs, to avoid overcounting.
  • H3 (Persistence): Attribute each sandwich to a bot — a single sender, a recurring sender pair, or a contract on EVM chains; the front-leg wallet on Solana — and keep only bots with at least 100 sandwiches, at least 60% of them profitable, and sandwich legs making up at least 25% of their swaps.

The H3 filter is the main departure from prior work. As the authors put it, when a bot makes a million swaps and fewer than 1% carry sandwich structure, those matches are more likely a by-product of its trading than genuine attacks. An ablation of all three heuristics is reported in Appendix C.

On Ethereum, the scope is restricted to transactions that never entered the public mempool, using mempool labeling and OFA victim attribution described in Appendix A. The authors are explicit about a coverage limitation: the OFA datasets provide positive evidence of observed exposure but cannot establish that a transaction did not appear elsewhere, so temporal OFA patterns are used to motivate mechanism-level analysis rather than as evidence of changing OFA usage.

Why This Matters

Impact on research. The paper reframes front-running protection as a property that must be evaluated end-to-end rather than assumed from a submission design. It shows that counting attacks measures attacker behavior, not the theoretical attack surface, and it gives the field a reusable detection methodology for non-tight sandwiches — the dominant shape once attackers lose visibility.

Real-world applications:

  • Wallet and RPC providers can use the exposure taxonomy to audit where transaction details leak, particularly the cross-OFA correlation risk between MEV-Share and MEVBlocker and the full-disclosure risk in whitelisted auctions.
  • OFA operators gain concrete evidence that whitelisting or obfuscation alone does not prevent sandwiching when a single searcher receives full transaction data, and that splitting disclosure across feeds can be recombined.
  • Chain and client developers on Tron and Solana learn that first-come-first-served ordering and local mempools reduce but do not eliminate latency-based front-running, and that application-level behavior can become the dominant leak once validator-level leaks are closed.
  • Traders and DEX users relying on "private" submission should recalibrate expectations: protected transactions can still be sandwiched, often from farther away and across block boundaries.

Industry relevance. Over $346M in net profit across a three-year window, with a single entity behind all of Ethereum's protected sandwiching, makes this directly relevant to builders, searchers, sequencers, and the teams operating the auctions and RPCs that users are told to trust.

Future Directions

  • Closing the unattributed gap. 37.3% of victims have no OFA attribution across the full period, and 61.0% of victims up to February 2025 remain unattributed. The MEVBlocker-only cluster around October–November 2024 and the MEV-Share-only spike in April 2026 are not explained by any identified mechanism, and the authors note an additional exposure source may be missing from their data.
  • Understanding why Arbitrum and Monad are clean. Their protections mirror Base's and Solana's respectively, yet no persistent sandwich bots appear. Whether this reflects stronger implementation, lower incentive, or simply less time (Monad launched 24 November 2025) is unresolved.
  • Separating incidental from targeted victims. Multi-victim sandwiches complicate attribution — 68.5% of Ethereum victims are sole targets, but 56.5% of unattributed victims occur in multi-victim sandwiches. Better methods for identifying which transaction an attacker actually targeted would sharpen every downstream claim.
  • Designing end-to-end defenses. The paper concludes that stronger end-to-end protection is needed but, in the available content, does not report a specific proposed design or an evaluation of countermeasures.

Target Audience

Researchers and practitioners working on MEV, blockchain security, and DEX infrastructure; protocol and client engineers designing transaction submission and ordering mechanisms; operators of RPCs, order-flow auctions, and sequencers; and analysts who need a rigorous methodology for distinguishing genuine sandwich attacks from high-frequency trading. Readers without blockchain background will follow the high-level argument but will need the background sections to interpret the structural statistics and the per-chain exposure analysis.

Note: the provided paper content is truncated partway through Section 6.1.5, so the detailed per-chain exposure analyses for Tron (Section 6.2), Base (Section 6.3), Solana (Section 6.4), and the reorg analysis (Section 6.1.6) are only available here at the summary level given in the abstract, introduction, and Section 5.

Authors’ abstract

Front-running has long plagued Ethereum's public mempool, earning it the nickname of a "dark forest", where predators lurk for profitable transactions. In response, Ethereum and other blockchain ecosystems increasingly rely on private RPCs and native protections to shield transactions from adversaries, which we refer to as protected order flow. Yet the effectiveness of these mechanisms in preventing front-running, and what trust assumptions they entail, remain poorly…

The full abstract on arXiv
Read the original paper