Correlated slashing in numbers: what it costs and what lowers it
A slashed validator loses more when more of the network is slashed in the same 36 days. The correlated part takes 3% of its balance at 1%, 15% at 5% and all of it at a third or more. On mainnet, 594 validators had been slashed in total by October 22, 2026, and two of the largest events on record involved 75 and 39 validators. For 10,000 ETH, a slashing of its validators alone costs about 25 ETH; with 5% of the network in the same window, about 1,516 ETH. Spreading a stake across operators with different setups lowers how much of it any one cause can reach.
By Develp
Ethereum penalizes a slashed validator more heavily when many validators are slashed around the same time. One validator slashed on its own loses about 0.2% of its balance. If 5% of all stake were slashed in the same window, each of those validators would lose about 15% to the correlated part alone. So the useful question about concentration is how much of the network shares a cause with your validators. The second is how much of your stake sits behind any one cause.
How a validator gets slashed
A validator is slashed for one of two offenses, and both can be proven on chain. It can propose two different blocks for the same slot. Or it can sign two attestations, its votes on the state of the chain, that contradict each other. The protocol can't tell an attack from a mistake, so it treats both the same way.
The cases on record include operators running the same validator keys on two machines at once. Upgrading Ethereum, a reference book on the protocol, traces every slashing it could examine to that cause, with one exception: a block proposer who attacked a relay. In February 2023, an Ethereum core developer counting 226 slashings said many had happened during failed system migrations.
Being offline is a different matter. A validator that misses its duties loses roughly what it would have earned, and it isn't slashed.
For a staker, this makes slashing an operational risk. It depends on how an operator runs its signing keys, including how it moves them between machines.
What a slashing costs
The cost of a slashing comes in four parts:
- The initial penalty is taken at once. Since the Pectra upgrade of May 2025, it is 1/4096 of the validator's effective balance (the stake the protocol counts), about 0.024%.
- The validator is forced out and can't withdraw for 8,192 epochs, about 36 days. Throughout that time it is penalized as if it were offline, which at today's total stake drains about 0.16% of its balance.
- About 18 days after the slashing comes the correlation penalty. It is three times the share of all stake slashed in the 36 days around the slashing, applied to the validator's balance and capped at the whole balance.
- If the chain stops finalizing while the validator is still out, the inactivity leak adds to the drain. The leak is a heavier penalty on inactive validators that starts after more than four epochs without finality.
The stake also earns nothing while it is out. To stake again, it goes through the withdrawal sweep and the entry queue. On October 26, 2026, those took up to 7.1 days and 22 days and 7 hours.
On its own, a slashing has a small and predictable cost. Nearly all of the uncertainty sits in the correlation penalty, which depends on how many other validators are slashed in the same window.
The correlation curve
The correlation penalty grows in a straight line with the share of stake slashed together, until it takes the whole balance at a third.
| Share of all stake slashed in the window | Correlation penalty, as a share of balance |
|---|---|
| One validator alone | close to zero |
| 0.1% | 0.3% |
| 1% | 3% |
| 5% | 15% |
| A third or more | 100% |
The initial penalty and the 36-day drain come on top of each row.
The shape is deliberate. Vitalik Buterin's design rationale for the beacon chain explains why. "A validator misbehaving is only really bad for the network if they misbehave at the same time as many other validators," he wrote. The scheme punishes a coordinated attack heavily and treats a single isolated failure lightly.
The penalty counts everyone slashed in the window, your own validators included. The number to know is the share of the network that shares a cause with your validators.
What has happened on mainnet
Large correlated slashing has never happened on mainnet. By October 22, 2026, 594 validators had been slashed since the beacon chain launched in December 2020, the last in August 2026. For scale, 823,287 validators were active the same day.
Two of the largest events on record involved 75 and 39 validators. In early February 2021, 75 validators run by one staking firm were slashed after the firm turned off a slashing protection database. On September 10, 2025, 39 validators were slashed after operator errors at two providers, during maintenance and a migration.
Measured the way the protocol measures it, the 2021 event was about 0.1% of the stake. The correlation penalty divides by the stake of active validators, which grew slowly after launch. The protocol admitted only four new validators per epoch, so no more than about 79,560 could have been active. Against that, the 75 validators' 2,400 ETH were at least 0.09% of the active stake.
At that share, the correlation penalty came to nothing. The rules then used a multiplier of one and counted the penalty in whole ETH, rounded down, so it stayed at zero below 1/32 of the stake. The firm put its penalty at around 18 ETH that week, which matches the initial penalties of 1/128 for each validator.
For comparison, 1% of today's stake is about 428,000 ETH, roughly 178 times the 2,400 ETH slashed in 2021.
Mass slashing has been seen on a test network. On the Medalla testnet in August 2020, nodes running one client had their clocks set hours ahead, and a large number of validators were slashed.
So far, slashings on mainnet have stayed far from the steep part of the curve. The reason to care is that shape, which would apply to every validator behind a shared cause.
What a large event would need
Many validators would have to sign conflicting messages within the same 36 days. One way this could happen is through something they share. That could be one operator's procedure, applied to a whole fleet during a migration or a failover. It could also be one piece of signing software or one hosting setup.
Today the stake sits with many operators. On October 22, 2026, Miga Labs tracked 773 staking entities with 100 or more validators each. On the software side, the largest consensus client runs on 49.38% of nodes in Miga Labs' count of October 26, 2026. That figure counts nodes, so it doesn't show how much stake one client signs for.
For a stake, the rate that one shared mistake would cost is set by how much of the network relies on the same operator, software and setup.
A worked example: 10,000 ETH
Take 10,000 ETH staked as 312 validators of 32 ETH, which hold 9,984 ETH. The table shows what a slashing of all 312 would cost at today's total stake.
| Validators slashed in the same window | ETH lost | Share of the stake |
|---|---|---|
| These 312 alone | about 25 | 0.25% |
| 1% of the network | about 318 | 3.2% |
| 5% of the network | about 1,516 | 15.2% |
| A third or more | 9,984 | 100% |
Source: protocol rules; network stake: Miga Labs, Oct 22, 2026
Every row counts the initial penalty (about 2.4 ETH), the 36-day drain (about 16 ETH) and the correlation penalty.
In every row the stake also earns nothing while it is out. At the network's rate over the last quarter, 2.61% a year, that comes to about 26 ETH.
Spreading the stake changes how much of it a single cause can reach. Suppose a first operator runs 5% of the network. A second runs 0.5%, with a setup that differs from the first in its clients, hosting providers and regions.
| Where the 9,984 ETH sits | Error across the first operator's fleet | Error across the second operator's fleet |
|---|---|---|
| All with the first operator | about 1,516 ETH | none |
| Split evenly between the two | about 758 ETH | about 84 ETH |
Source: protocol rules; network stake: Miga Labs, Oct 22, 2026
The operator sizes are illustrative, and the two errors are assumed to be independent. Each figure counts the same three parts as the first table.
Two numbers set the exposure. The share of the network behind a cause sets the rate, and the share of your stake behind that cause sets how much the rate applies to. Spreading a stake across operators lowers the second. Choosing setups that differ from where most of the network sits keeps the first small.
What spreading doesn't do
Spreading doesn't lower the rate in a network-wide event, because the penalty counts every validator slashed in the window. It doesn't replace slashing protection either: each signer keeps a record of what it has signed and refuses to sign anything that conflicts. When keys move between clients, EIP-3076 defines a format for carrying that record across. Distributed validator technology (DVT) splits a key so that no single machine can sign alone, but operator error can still reach it.
Correlated downtime is a separate mechanism with a separate penalty. On December 4, 2025, a bug in one consensus client cut network participation to as low as 75% across 42 epochs. Finality held, less than nine percentage points above the two-thirds threshold, so the inactivity leak never started.
The cost was about 382 ETH in attestation rewards that validators missed. A client bug can also lead to slashing. Ethereum.org notes that a critical bug in a client run by two thirds of the network could split the chain and finalize the wrong side. Validators on that side would then face slashing, or a slow and costly exit and re-entry, to rejoin the correct chain. Client diversity is about keeping every client below that line, and Develp's team builds Nimbus, one of Ethereum's consensus clients.
How the rules have changed
| Upgrade | Date | Initial penalty | Correlation multiplier |
|---|---|---|---|
| Beacon chain launch | Dec 1, 2020 | 1/128 | 1 |
| Altair | Oct 27, 2021 | 1/64 | 2 |
| Bellatrix | Sep 6, 2022 | 1/32 | 3 |
| Pectra | May 7, 2025 | 1/4096 | 3 |
Source: Ethereum consensus specs; dates: ethereum.org
Bellatrix raised both values to the levels its specification describes as final. Pectra then cut the initial penalty. EIP-7251 let a validator hold up to 2,048 ETH, and a fixed share of a larger balance would have put more at risk on large, consolidated validators. The correlation multiplier stayed at three.
The next upgrade, Glamsterdam, is set to include EIP-8045, under which slashed validators would no longer be picked to propose blocks. Today they can still be picked, but their blocks are invalid and the slot stays empty. Glamsterdam has no mainnet date yet.
With the fixed part of the penalty now small, nearly all of the risk sits in the correlated part.
Questions to take to any provider
If you are reviewing an operator, it's worth asking:
- How do you keep a key from signing in two places during a migration or a failover?
- How is signing history carried over when keys move between machines or clients?
- What share of the network's stake do your validators represent?
- How are your validators spread across clients, hosting providers and regions, and what is the largest share of our stake one cause could reach?
- Have any of your validators been slashed, and what changed afterward?
These questions apply to any operator, Develp included. If you'd like to go through them for your own stake, talk to our team.
Questions
What is correlated slashing?
It is the part of Ethereum's slashing penalty that grows with the share of all stake slashed in the same 36-day window. A validator slashed alone pays close to nothing in this part, while at a third of the stake or more it loses its whole balance.
How much does a single slashing cost?
About 0.2% of a validator's balance at today's total stake: a small initial penalty, then about 36 days of being penalized as if offline. The stake also earns nothing while it is out, and it waits in the queues to stake again.
Has mass slashing happened on Ethereum mainnet?
No. By October 22, 2026, 594 validators had been slashed in total, and two of the largest events on record involved 75 and 39 validators. Mass slashing has been seen on a test network, Medalla, in 2020.
Does spreading stake across operators lower the penalty?
It lowers how much of your stake one cause can reach. It doesn't lower the rate in a network-wide event, because the penalty counts every validator slashed in the window.
Can a client bug get validators slashed?
A bug that takes validators offline costs missed rewards, and more if finality stops. One route to slashing is a critical bug in a client run by two thirds of the network that finalizes the wrong chain. Validators on that chain would face slashing to rejoin the correct one.
Sources
- Bellatrix beacon chain, Ethereum consensus specs
- EIP-7251: Increase the MAX_EFFECTIVE_BALANCE, Ethereum Improvement Proposals
- EIP-8045: Exclude Slashed Validators from Proposing, Ethereum Improvement Proposals
- Serenity design rationale, Vitalik Buterin
- Client diversity, ethereum.org, 2026-02-23
- Upgrading Ethereum, 2.8.7 Slashing, Ben Edgington
- Slashed Validators, Miga Labs, 2026-10-22
- Ethereum Staking Entities, Miga Labs, 2026-10-22
- Seventy-five Eth2 validators got slashed this week, The Block, 2021-02-04
- Ethereum hit by rare mass slashing: 39 validators penalized, Unlock Blockchain, 2025-09-11
- Only 0.04% of Ethereum validators have been slashed since 2020, says core dev, Cointelegraph, 2023-02-23
- What's New in Eth2, 22 August 2020, Ben Edgington, 2020-08-22
- Validator queue, validatorqueue.com, 2026-10-26
- Glamsterdam Testnet Announcement, Ethereum Foundation, 2026-09-17
- Mainnet postmortems, Offchain Labs
- Client diversity, clientdiversity.org (Miga Labs data), 2026-10-26
- Ethereum sees 25% validation drop post-Fusaka as Prysm bug nears finality loss, Cointelegraph, 2025-12-05
- EIP-3076: Slashing Protection Interchange Format, Ethereum Improvement Proposals
- The history of Ethereum, ethereum.org
