Why it exists
A proof-of-stake chain needs a cost for cheating. In proof of work the cost is the electricity already spent, which cannot be recovered. In proof of stake the cost is the validator's own stake, which the protocol can destroy. ethereum.org describes slashing as a mechanism where some or all of a validator's staked ETH can be destroyed if the validator acts dishonestly.
What triggers it
Slashing is reserved for behaviour that threatens consensus and can be proven on chain, such as signing two different blocks for the same slot or making contradictory votes. Merely being offline is penalised too, but more lightly, through small deductions rather than slashing. This distinction matters: ordinary downtime is a small cost, while attacking the rules is expensive.
How severe it can be
The size of the penalty depends on the situation. ethereum.org explains that the amount varies with how many validators are slashed at the same time, from a minor penalty for an isolated mistake to losing the full stake when many validators misbehave together. That correlation penalty is meant to make coordinated attacks very costly.
What it means for stakers
If you run your own validator, avoid running the same keys on two machines, because duplicate signing is the most common cause of accidental slashing. If you delegate or use a pool, understand how the provider handles slashing and whether you share the loss. Good staking providers publish their practices and often hold insurance or buffers, but these are not guarantees. Slashing is rare in practice, but it is the reason staking is not risk-free.