PRIVATE pools without sandwiches? No! — that doesn’t exist!

Hello, Marinade community! This time we’re coming to you with new research — we’re sure you’ll find it interesting!

Old Problems in a New Wrapper

We received reports from fellow validators: they were approached with offers to join a so-called “private mempool without sandwiches” with promises of “security and high yield.” This sounded too good to be true. Together with @w3e_uk, we decided to validate these claims using on-chain metrics.

Summary

We observed recurring sandwich attacks originating from a specific program:
Program ID: inf69quFVZyuHEsrUXq3APtYLr4iqsNiQdCh5ArGcUp

Within a single slot, the front → victim → back pattern is clearly visible, often with splitting (multiple buys/sells around the same victim).

Based on slot distribution, the contract appears to be sourcing flow via a private mempool connected to a subset of validators.

From our initial sample, a significant portion of these validators hold delegated stake from @MarinadeFinance, @jito_sol, and @edgevana.

What We Measured

We collected transactions from the identified program and analyzed the blocks (slots) where it appeared. For each such block we:

  • Reconstructed the transaction ordering

  • Identified front-run → victim → back-run frames on the same token within a single slot

  • Allowed for splitting: multiple entries before the victim and multiple exits after

For each sandwich event we recorded:

  • Transaction indices within the block (positions)

  • Signatures of front/victim/back transactions

  • Number of entries and exits executed by the attacker

On-Chain Evidence

The sandwich pattern (front → victim → back) is consistently present, frequently with splitting. The contract often fragments both the “entry” and “exit” into multiple smaller transactions within the same block. This technique:

  • Obfuscates traces for superficial detection systems

  • Creates a false sense of trust in the private mempool, since monitoring tools like sandwiched.me, @0xGhostLogs, and sandwich finder @0x7cf are not yet capable of reliably detecting such fragmented sandwiches

A clear example of a sandwich attack:

Buy 1:

3y185hmf8ZpFiD2gsqmxr5Xjerri7Y2noGdjBVPLZUdF5wsz7DyETjxZkXRryj9s56FZNwQCdLCuVw81y9DV89M1

Buy 2:

3mLDZaa4nWgsJdcEc7qmJtGuKDLJvULgMiMqQoAgCPx3Sj9THfPhzh4EKam7tTcTDyBxyGsmjqBufuqGarbGmXcQ

Transfer token to other wallet:

64htjTV51R5AhUTi6Yof5m3jzt62LasVzUFEkBXAb252d6teLwwQQe9rUVdXVFusr1vHEE8HR8gzbMGNJeKfzwAe

Victim:

3ngZ6P8rJJYRQERvMESwKozBPKDgf23sJiEDvvbZQxFyXQdnp7n5CZkY7vn6bswQib1wb9pdg8FcjtiX8S53B6dz

Sell all tokens:

4S2PJhoKWjzz1PqcsHYiyjPHByYLqAFK7TKAXkqFpXW56KicjWsfQX3PjJstw1Uvs4w8wwMAvByzv75udVJaDEam

Validators Processing Contract Flow

At the link below we provide a list of validator identity keys whose blocks contained transactions from the program (based on our dataset). For transparency, we include their activated stake (SOL) and tx_count (the number of times the program appeared in their slots within the analyzed range).

Conclusion:

Our sample shows significant delegations (including from Marinade, JITO, Edgevana, and others) to these validators. In practice, however, their blocks consistently carry traffic from a program exhibiting clear sandwich-bot behavior. This directly contradicts the claims of a “private channel without sandwiches.”

We will continue monitoring and publish further updates.

We hope the Marinade team and the teams behind sandwich attack analysis services will take note of this new sandwiching technique and promptly adjust their detection algorithms accordingly.

Thank you for this research. Do you have a dashboard and repearable methodology to detect these? If so, I’d be keen on exploring further.

Good day. No, we haven’t worked in this direction yet. The purpose of our post is to signal to services that work with sandwich formation, as well as stakepools, so that they quickly adapt to new sandwich hiding schemes.

We have python code for analyzing transactions and identifying a new type of sandwich, if you are interested, you can share it, we don’t mind)

Hey everyone,

A quick update on the new so-called “private pool without sandwiches”.

Things are moving fast: the number of connected validators has jumped from 17 to 31 (almost 2x growth), and the connected stake has skyrocketed from 1,318,031 SOL to 3,498,194 SOL (2.6x increase).

Almost all validators in the pool are heavily backed by Marinade stakes.
We hope this time the team will react quickly to address the issue.

Stay tuned — we’re actively monitoring the mempool dynamics and will keep you posted!

P.S. I’ll attach a link to a Google Sheets document with each post, and every new update will be on a separate tab. The current relevant tab is 19.08.2025.

Hello everyone.

It’s quite unfortunate that the Marinade team remains completely silent on this issue and hasn’t taken any action to revoke bonds or unstake from validators who are actively running sandwich attacks — cleverly hiding behind position splitting to make it look “clean.”

But unlike them, we do care. That’s why we keep publishing the stats.

As of August 22, the total number of validators increased by one and now stands at 33. The total stake has grown to 3,523,318 SOL (up from 3,498,194 SOL, a delta of +25,123 SOL).

For epoch 837, only 24 validators from our list receive Marinade stake, with a combined total of 2,155,774 SOL delegated. At the moment, Marinade controls roughly 9.6 million SOL — and over 20% of that stake is being used to execute sandwich attacks.

Sad? Absolutely.
What’s even sadder is that Marinade’s shiny new “reputation system” doesn’t change the reality: Marinade reacts painfully slowly to new challenges in the ecosystem.

Link for out table - Table - Google Sheets

Hello everyone,

We’re publishing an updated report on this topic.
As of now, the number of validators has decreased to 18 (down from 32 in the previous report).
The total stake has dropped from 3.5M SOL to 2.6M SOL. Out of the 2.6M SOL, nearly 2M SOL comes from Marinade’s delegation.

Unfortunately, the Marinade team still hasn’t provided any feedback in this thread — which is disappointing.

See you in the next update.

Table llink is here - Table - Google Sheets