MEV explained: sandwiches, arbitrage, liquidations and protection

Ordering is worth money

Maximal extractable value is the profit available to whoever decides which transactions go into a block and in what order. The term started life as "miner extractable value" in the 2019 paper Flash Boys 2.0, which documented bots bidding up gas in priority gas auctions to jump the queue on early DEXs. The M is now read as maximal, which suits proof of stake Ethereum, where validators rather than miners produce blocks. The economics are unchanged: if a pending transaction will move a price, someone can profit by standing next to it.

On Ethereum the work is split between specialists. Searchers run bots that scan pending transactions and new blocks for opportunities and package their trades as bundles. Builders assemble the most valuable block they can from bundles and ordinary transactions. Relays pass the winning block to the proposer, the validator whose slot it is, through MEV-Boost. In the week to late September 2026, relayscan.io recorded relay delivered blocks for roughly nine in ten Ethereum slots, and three builders produced about 92% of those blocks. The full supply chain, and the plan to move this auction into the protocol itself, is the subject of the Expert guide to proposer builder separation.

Competition matters to you in one specific way. Searchers bid against each other for position, so on contested opportunities much of the value ends up paid to builders and proposers rather than kept by the bot. That does not help the user who was extracted from. It does mean that the size of the opportunity your transaction creates, not the cleverness of any one bot, sets what you lose.

The chain also runs on trust. On April 3, 2023, a validator exploited a flaw in widely used relay software: it signed an invalid block header, the relay revealed the block's contents anyway, and the validator rebuilt the block to break apart sandwich bots' bundles, taking about 20 million dollars from them according to the Flashbots post-mortem. Relays were patched to reveal a block only after publishing it successfully. The hunters, it turned out, were also prey.

A field guide to MEV

The useful split is not good against evil but who bears the cost. Arbitrage and liquidations extract value, yet they perform a service the system needs, and the people who pay for them, liquidity providers and borrowers, accepted that role when they deposited. A sandwich performs no service. It exists only because a trade was visible before it landed and its minimum output was loose.

Arbitrage and backruns: the benign majority

Take a constant product pool holding 1,000 ETH and 3,000,000 USDC with a 0.3% fee, so ETH is 3,000 in the pool and, for the example, 3,000 everywhere else. Someone buys ETH with 30,000 USDC. With the swap formula from AMM math, they receive 9.8716 ETH, an average of 3,039.03, and leave the pool quoting 3,060.21.

A backrunner now buys ETH at 3,000 elsewhere and sells about 8.41 ETH into the pool until the pool's price, net of the fee, matches the outside market. That makes about 214 USDC before gas and leaves the pool near 3,009, the edge of the band a 0.3% fee allows. Nobody was front run. The swapper got exactly the quote they saw. The 214 dollars came out of the pool's liquidity providers, who sold ETH too cheaply for one block.

That last sentence is the whole economics of providing liquidity. Every arbitrage profit is a loss to LPs compared with trading at the market price, a cost the research literature calls loss versus rebalancing. The guide to LP as a strategy measures it.

Liquidations: MEV that keeps lenders solvent

A lending market cannot sell your collateral itself, so it pays outsiders to do it. When a loan's health factor falls below 1, anyone may repay part of the debt and receive collateral worth more than they repaid. Illustrative numbers, since every market sets its own parameters:

Liquidators race for these, and the race often starts when an oracle update lands: the price update is the transaction that makes the loan liquidatable, so bots try to land directly behind it in the same block. The bonus is the borrower's cost, set high enough that someone will always show up in a crash. How it is chosen, and what happens when nobody shows up, is covered in inside a lending protocol and the guide to oracles.

The sandwich, step by step

1. The bot sees your swap

   Your transaction sits in the public mempool with its full calldata, including the minimum output your slippage tolerance produced.

2. It buys first

   The front run buys the token you are buying and pushes its price up. The bot sizes it so your trade still clears at exactly your minimum, because a victim whose trade reverts pays the bot nothing.

3. Your swap executes at the worse price

   You receive your minimum, and your own buy pushes the price further up.

4. It sells straight after

   The back run sells what the bot bought, at the higher price your trade created. All three transactions travel as one bundle, so the bot takes no market risk.

Run it on the same pool and the same 30,000 USDC buy, quoted at 9.8716 ETH. The bot solves for the largest front run your tolerance allows:

Three things fall out of the numbers. Your loss is your tolerance times the trade, because the attack is sized to hit your minimum. The bot keeps about 70% of what you lose; most of the rest is the 0.3% fee it pays the pool twice. And the bot's profit is gross: gas and its bid to the builder come out of it, which is why a 0.1% tolerance on a small trade is rarely worth attacking and a 5% tolerance on almost any trade is.

The extreme case happened on March 12, 2025. A swap of about 220,764 USDC into USDT on a Uniswap v3 stablecoin pool returned about 5,271 USDT. According to reports at the time, a bot traded the pool to the edge of its liquidity just before the swap and reversed the trade just after, keeping around 8,000 dollars and paying roughly 200,000 dollars to the block builder for its position. Researchers questioned whether the sender was an ordinary trader at all, but the mechanism holds no mystery: a swap whose minimum output is close to zero is an open offer.

Protection on Ethereum and other EVM chains

Every protection removes one of the two things a sandwich needs: visibility before inclusion, or a loose minimum.

Details from the product documentation, as of September 2026: Flashbots Protect has a fast mode that shares your transaction with all registered builders to speed inclusion, and it warns that switching RPCs in some wallets while a protected transaction is pending can rebroadcast it to the public mempool. MEV Blocker, created by CoW Protocol with Beaver Build and Agnostic Relay and operated since January 2026 by Special Mechanisms Group, part of Consensys, offers several endpoints that trade revert protection against rebates. Owners and defaults change; the questions to ask do not. Who sees my transaction, who may act on it, and where does any refund go?

Layer 2s

Most rollups run a single sequencer, which makes ordering a question of that sequencer's policy. Arbitrum is a documented example. It ordered transactions first come first served, which its documentation says rewarded latency races and spam, and it now runs Timeboost on Arbitrum One and Nova: an auction sells control of an express lane in 60 second rounds, and everyone else's transactions wait an extra 200 milliseconds by default. The docs state the winner gets only a time advantage, not the ability to see or reorder other transactions, and the mempool stays private. On chains like this the sandwich from a public mempool largely disappears, and MEV shows up as arbitrage competition instead.

Solana: no public mempool, different MEV

Solana has no mempool in the Ethereum sense. RPC nodes forward transactions directly to the validators scheduled to lead the next slots, and a transaction that is not processed expires after roughly 150 blocks, about a minute. There is no global waiting room for bots to watch. That removes the easy version of the sandwich, not the incentive.

Much of Solana's MEV runs through Jito. Validators running the Jito-Solana client connect to Jito's block engine, which accepts bundles: up to five transactions that execute in order, all or nothing, within one slot. Bundles compete on tips, a SOL transfer to one of eight tip accounts with a 1,000 lamport minimum, in auctions the engine runs every 50 milliseconds among bundles that touch the same accounts. Tips flow to validators and their stakers. Ordinary priority is a separate channel: a priority fee, usually priced in micro lamports per compute unit, goes entirely to the validator, while half of the 5,000 lamport base fee per signature is burned.

Sandwiching on Solana moved into private channels instead. Jito ran a mempool service until March 8, 2024, when it shut the service down because searchers were using it to sandwich users. Private mempools run by others took its place, and in June 2024 the Solana Foundation removed a group of validator operators from its delegation program for taking part in them. Two newer defenses are worth knowing, both described in Jito's documentation:

• jitodontfront. Add any valid public key beginning with jitodontfront to one of your transaction's instructions (the account need not exist onchain) and Jito's block engine rejects any bundle that places your transaction anywhere but first. Jito calls this a mitigation that applies only to its own block engine, not a complete defense.
• BAM. Jito's Block Assembly Marketplace schedules transactions inside trusted execution environments, so pending transactions stay encrypted until execution and the ordering is attested. It went live in early mainnet form on September 25, 2025, and as of mid 2026 a minority but growing share of stake ran it.

For users the lesson is the same as on Ethereum: slippage is the budget an attacker works with. Routers such as Jupiter now estimate it automatically from market conditions, which beats a fixed wide number. The Solana runtime guide covers the fee market underneath.

Checking whether you were hit

1. Compare the fill with your minimum

   Read the token transfers on the transaction. Receiving almost exactly your minimum is a fingerprint of a sandwich, because attacks are sized to it and ordinary drift rarely lands on it to the last unit.

2. Look at the neighbors

   On an EVM explorer, open the block and find the transactions listed immediately before and after yours. A buy of your token just before and a sell just after, from the same contract, is the pattern.

3. On Solana, look for the bundle

   The Jito explorer shows bundles and their tips. A bundle holding a buy, your swap and a sell tells the same story.

4. Fix the cause

   Tighten the tolerance, route through a protected path, split a large trade, or use an intent based order. The guide to reading transactions covers the explorer work in detail.

For the basics of slippage and minimum output, see getting a good fill. The short glossary entries on MEV and sandwich attacks are handy to share with someone newer.

Common questions

Where to go next